Choosing the right machine learning algorithm for business success

Algorithm

Machine learning can be overwhelming with its variety of tasks. Most tasks can be solved with a few ML algorithms. You need to be aware of which algorithms to select, when to apply them, what parameters to take into consideration, and how to test them. This guide was crafted to provide you with a straightforward and actionable solution to the issue at hand.

Choosing the right model for production is important. Not all algorithms should be used for every problem, even though we have many performance metrics to evaluate models. It’s time-consuming and labor-intensive. Selecting the right algorithm is important.

Data governance strategy and features are important when choosing a machine learning algorithm. ML models work with various datasets that differ in data points and features, depending on the use case. Model selection depends on how the model handles different dataset sizes.

Neural networks work well with big data and many features. SVM can handle limited features. Consider data size and feature count when choosing an algorithm.

What is a machine learning algorithm?

Let’s start with the basics in case you’re unsure about what this is and why you might need it. We’ll discuss machine learning and its algorithms. Skip to the guide on choosing ML algorithms if you already know this.

Learning algorithms examine data to recognize patterns and formulate hypotheses. ML algorithms are computers that have been trained. These are the different classifications of ML algorithms, which can be broken down into three and a half groups.

Humanity generates increasing amounts of data daily. It comes from many places, like business data, social media, IoT sensors, and more. ML algorithms turn data into useful information for automation, personalization, and complex predictions.

ML algorithms specialize in specific tasks based on the data features and project requirements. Principal categories of machine learning algorithms, along with illustrative examples for various kinds of jobs.

Steps to choosing the right machine learning algorithm

Define the problem

Define the problem you want to solve with machine learning. Business objective, target audience, current situation, desired outcome? How to measure success and evaluate performance? Assumptions, risks, and constraints impacting the problem? Define the problem, scope the project, identify data sources and features, and choose machine learning techniques.

Understand the data
Next, understand your data and the data needed for your machine learning project. Data is crucial for machine learning. It should be relevant, reliable, and representative of your problem domain. Clean and preprocess the data for machine learning. Perform descriptive and inferential stats, visualize data, check for missing values, outliers, and anomalies. Consider the ethical and legal implications of data usage and sharing.

Build the model
Build the model to learn from data and make predictions or recommendations. Choose the correct algorithm, architecture, and parameters for your machine learning task. Split data into training, validation, and test sets. Apply cross-validation, regularization, and other techniques to avoid overfitting or underfitting. Use libraries and frameworks to simplify and speed up model building.

Evaluate the model
Evaluate the model’s performance on the test set and new data. Choose the right metrics for your machine learning task and business goal, like accuracy, precision, recall, F1-score, ROC curve, or others. Compare models with baseline models, analyze errors, biases, and uncertainties. Consider the interpretability, explainability, and fairness of the model and its impact on stakeholder trust and satisfaction.

Deploy the model
Deploy the model into production for customer use. Ensure the model is scalable, robust, secure, maintainable, and can handle real-world data and scenarios. Monitor and update the model regularly. Collect feedback and metrics on its impact on business goals and outcomes. Use tools and platforms for ML model deployment and management.

Communicate the results
Step six: Communicate results to stakeholders. Tell a story that shows the value and benefits of your machine learning solution, and how it solves the problem and meets the business goals. Present data, model, and evaluation clearly and engagingly with visualizations, charts, tables, and other elements. Address limitations, challenges, and improvement opportunities for your machine learning project.

Types of machine learning algorithms

  • Supervised learning
  • Unsupervised learning
  • Reinforcement learning

Supervised learning
Supervised learning involves human guidance to teach the algorithm. Supervised learning needs known outputs and labeled training data. If output is real, it’s regression. If the output is from a limited number of unordered values, it’s classification.

Unsupervised learning
Unsupervised machine learning is aligned with true artificial intelligence, as it allows computers to learn complex processes and patterns without human guidance. Less info about objects, train set unlabeled. Objects can be grouped based on similarities. Some objects can be anomalies.

Reinforcement learning
Reinforcement learning is about algorithms that learn to achieve goals or maximize outcomes over many steps. Maximize points won in a game. Supervised learning uses labeled training data, while reinforcement learning does not have labeled data. Without training data, it learns from experience.

Commonly used machine learning algorithms

1-Linear Regression
Linear regression summarizes relationships between two continuous variables. X is the independent variable. Y is the dependent variable. Linear regression uses X to predict y, while multiple regression uses multiple X variables to predict y based on a loss function like MSE or MAE. Use a regression algorithm to predict future values of a running process. This algorithm works well with thousands of features, like bags of words or n-grams in natural language processing. Complex algorithms overfit features, while linear regression is decent. Unstable if features are redundant.

Choosing the right machine learning algorithm for business success

2-Logistic Regression
Don’t mix up classification algorithms with regression methods. Logistic regression does binary classification with binary label outputs. Logistic regression is a special case of linear regression for categorical output variables, using a log of odds as the dependent variable. What’s great about logistic regression? It combines features and applies a sigmoid function, like a small neural network.

Choosing the right machine learning algorithm for business success

4-KNN

  • They have different goals. K-nearest neighbors is a classification algorithm in supervised learning. K-means is a clustering algorithm in unsupervised learning.
  • To assign positions to football players in a new dataset without positions but with measurements, we can use K-nearest neighbors.
  • For football player grouping based on similarity, we can use K-means. The K in each case means different things.
  • K-nearest neighbors uses K to represent the number of neighbors involved in determining a new player’s position. Example: K=5. To assign a position to a new football player, we select the five players with measurements closest to the new player and have them vote on the position.
  • K-means uses K to determine the number of clusters. If K=7, I’ll have seven clusters of football players after running the algorithm on my dataset. Two algorithms with different purposes, but both using K, can be confusing.

Tips for choosing the right algorithm for your business

  1. Identify your goals: Identify your goals to choose the right algorithm. What’s your goal with machine learning? What do you want to do with the data? Understanding goals helps narrow options and choose suitable algorithms.
  2. Consider the type and quality of your data: Consider data type and quality when choosing an algorithm. Some algorithms need labels, others don’t. Consider data size, complexity, and missing/corrupted points.
  3. Evaluate the available algorithms: Evaluate algorithms once you understand your goals and data. Online resources can help you compare algorithms and their pros and cons.
  4. Test and evaluate: Test and evaluate algorithms to find the best performer for your data. Use cross-validation and holdout sets to test algorithm accuracy and performance and compare results.
  5. Consider scalability: Consider algorithm scalability. Choose an algorithm that can handle large data or future expansion.

Conclusion

Selecting the right machine learning algorithm for your business can greatly impact the success of your project. Follow these blog post tips to choose an algorithm that meets your needs and goals. The right algorithm can maximize machine learning and business success.

This guide simplifies machine learning algorithms for decision-making. Choosing the right model is crucial as not all algorithms are suitable for every problem. Data and features are important for algorithm selection. NNs are good for big data and many features, SVM is good for limited features.

Ex-Nvidia & Ola Exec Launches RagaAI for testing and fixing AI

RagaAI, an AI-focused startup, has come out of stealth mode and has successfully closed a $4.7m seed funding round. Pi Ventures spearheaded the funding round, joined by international investors such as Anorak Ventures, TenOneTen Ventures, Arka Ventures, Mana Ventures, and Exfinity Venture Partners.

RagaAI addresses the need to ensure performance, safety and reliability of AI models, by providing an automated and comprehensive AI testing platform for companies. RagaAI is backed by advisors from Amazon, Google, Meta, Microsoft and NVIDIA.

RagaAI DNA

RagaAI DNA which is the foundational model of RagaAI uses automation to detect issues, diagnose and fix them. Offering over 300 different tests, the model is able to identify issues such as data drift, edge case detention, poor data labelling, bias in data, and many more. Furthermore, it is a multimodal platform that supports LLMs, images/videos, 3D, audio, NLP and structured data. It reduces 90% of the risks while accelerating AI development by more than 3x.

Tech Pioneer

Coming from a rich technology background of computer vision and machine learning, RagaAI was founded by Gaurav Agarwal in January 2022. He has previously worked with Texas Instruments and moved on to head mobility business at Ola and computing giant NVIDIA.

“At Ola & NVIDIA, I saw the significant consequences of AI failures due to lack of comprehensive testing. Our Foundation Models “RagaAI DNA” is already solving this problem across large fortune 500 companies,” said Gaurav Aggarwal, CEO and founder of RagaAI.

The founding team at RagaAI has a collective AI expertise of over 50 years. The company has already provided solutions for companies in various sectors including ecommerce, automotive, and others, with multiple use cases.

RagaAI’s funding round will be used to advance research and development, with a focus on improving AI testing tools.

“Driven by their patent-pending drift detection technology, RagaAI, an AI testing platform, is well-suited to solve these massive problems for the AI deployments globally. At pi Ventures, we believe in backing founders who can create disruptive solutions for global impact. In our view, Gaurav and his stellar team at Raga are fulfilling that goal in a big way. We are pleased to be associated with them,” said Manish Singhal, founding partner of pi Ventures that spearheaded the funding round.

The post Ex-Nvidia & Ola Exec Launches RagaAI for testing and fixing AI appeared first on Analytics India Magazine.

Delayed Approval of AI-Medical Devices Hampers Innovation

While a number of AI-enabled devices are out in the market, and significant developments in drugs and medical devices are also happening, the task of releasing them to the market is no easy feat. Governed by regulatory bodies, any drug or medical device undergoes stringent checks and trials before it hits the market.

Recently, the US FDA (Food and Drug Administration) cleared an AI device last week that can assist in detecting all types of skin cancer. The AI-powered device DermaSensor is reported to aid physicians in detecting skin cancer, with its pivotal trial demonstrating a sensitivity of 96% and a 97% probability of accurately identifying a skin lesion as benign. While the product was only recently approved, the approval process for DermaSensor began eight years ago.

Obstacles for Path To Launch

The U.S. Food and Drug Administration – FDA, which is a federal agency within the Department of Health and Human Services, plays a vital role in regulating the development and marketing of any new medical drug or device, however, the process to get approval from the department is lengthy.

The US has one of the most stringent regulations for medical drugs and devices in the world. The average time from FDA application to approval of drugs is 12 years. Furthermore, the projected average cost of bringing a new drug from concept to market surpasses $1 billion.

The only biggest exception to the lengthy problem, was witnessed during the pandemic. Pfizer and BioNTech were granted Fast Track Designation for mRNA-based vaccine development, thereby expediting the release process to less than a year.

Significant advancements are also being made to smoothen out approvals and upgrades for AI devices. Last year, the FDA proposed a plan that will allow developers of AI-enabled medical devices to automatically update their products that are already being used in the clinic.

“While dozens of companies have attempted to address this problem in recent decades, we are honoured to be the first device cleared by the FDA that provides PCPs (Primary Care Physicians) with an automated tool for evaluation of suspicious lesions,” said Cody Simmons, co-founder and CEO of DermaSensor.

In India, the Central Drugs Standard Control Organisation (CDSCO) takes around six to nine months for completing any medical device registration, only if a technical presentation or Subject Expert Committee (SEC) review is not required. The registration does not expire as long as the maintenance fee is paid every five years. As of now, there is no separate provision for AI-enabled devices.

AI-enabled device for detecting skin cancer

AI Assistance Might Transform Oncology

While the discussion on AI replacing doctors is far-fetched, the role of AI in diagnosis and assisting doctors are finding large use cases. Traditional methods of skin cancer detection involve dermoscopy where a medical professional physically examines a patient’s skin and lesions. However, this approach has room for human error. The diagnosis accuracy for such screening techniques is estimated at 75-84%. DermaSensor delivers a 97% of accuracy for identifying lesions as benign. The device can detect all three types of skin cancers-melanoma, basal cell carcinoma and squamous cell carcinoma.

“We are entering the golden age of predictive and generative artificial intelligence in healthcare, and these capabilities are being paired with novel types of technology, like spectroscopy and genetic sequencing, to optimise disease detection and care,” said Simmons.

When brought in contact with the skin, DermaSensor emits light, and then captures the wavelengths of light reflecting off cellular structures beneath the skin’s surface. It then utilises specific algorithms to analyse the reflected light and detect the presence of skin cancer. The algorithms are trained on data of more than 4000 malignant and benign lesions.

A number of deep learning-based methods have been proposed to assist dermatologists in accurate diagnosis of skin cancer.

Last week, MIT released two AI-powered models, PRISM neural network (PrismNN) and logistic regression (PrismLR), that helps with early detection of pancreatic cancer. The model is reported to identify 35% of cancer cases, while conventional screening methods only result in a 10% identification rate.

As the regulatory bodies aim to ensure a safe deployment of AI-enabled devices, delays in approval are likely to persist.

The post Delayed Approval of AI-Medical Devices Hampers Innovation appeared first on Analytics India Magazine.

Oracle’s Generative AI Madness Begins 

Oracle is integrating generative AI capabilities to the entire OCI’s tech stack. In an exclusive interview with AIM, Vinod Mamtani, vice president and general manager of Generative AI Services at Oracle Cloud Infrastructure, revealed Oracle’s plans which include the general availability of generative AI services, the launch of generative AI agents, and the introduction of Data Science AI Quick Actions.

“With a common architecture for generative AI that is being integrated across the Oracle ecosystem, the company is bringing generative AI to where exabytes of customer data already reside, both in cloud data centres and on-premises environments,” said Ritu Jyoti, Group VP, IDC.

Notably the enterprise friendly cloud service provider last year added generative AI services to several applications in the Oracle application suite which included HCM, SCM, CX, ERP, NetSuite, and a bunch of other industry vertical models.

Integrates Llama 2

The recent surge in the popularity of Llama 2 hasn’t escaped OCI’s attention and it will now host Llama 2 with 70 billion parameters. Interestingly, this is the first time OCI is hosting models other than Cohere.

When asked what made OCI turn to Llama 2, Mamtani said, “Regarding Llama 2, we noticed the interest and uptake from developers, which is why we thought it could be quite useful for our customers.” He further added, “We want to add Llama 2 7B within six weeks, and then we have Cohere’s Command, Summarise, and Embed models”.

Moreover, in addition to hosting Llama 2, OCI will now host Cohere’s 52 billion embedded models. “If you look at the leaderboard, Cohere’s embedded models are ranked very high. It supports both English and multilingual language embeddings,” said Mamatani, explaining that embeddings generated will be in the same space, irrespective of the language used.

The OCI Generative AI service also offers flexible fine-tuning, available for Cohere’s Command 52/6B models through vanilla and TPU fine-tuning. Moreover, in order to make it easier for customers to build their AI applications, Oracle has integrated LangChain.

Bets on AI Agents

Enterprises have a vast corpus of documents. Since LLMs are trained on public datasets, they are unaware of all the information in these documents. To address this, OCI has introduced Retrieval Augmented Generation Agents to its OpenSearch. Users can now simply attach their own documents to it and start chatting with it in natural language.

“It’s going to provide responses that are grounded, reducing hallucination. We are plugging in support for OpenSearch. Users can now transparently access diverse enterprise data sets through natural language without the need for specialist skills or to know the data’s format or storage location,” explained Mamtani.

Apart from the RAG agent, Oracle plans to introduce new agents. Upcoming releases will support a broader range of data search and aggregation tools and provide access to Oracle Database 23c with AI Vector Search and MySQL HeatWave with Vector Store.

Oracle will also deliver prebuilt agent actions across its suite of SaaS applications, including Oracle Fusion Cloud Applications Suite, Oracle NetSuite, and industry applications such as Oracle Health.

“There are these complex workflows that exist in our SaaS suite of apps. So, we want to automate and simplify that. And we will think of building agents that are very specialised for those applications.” said Mamtani.

Besides, Mamtani said that Oracle also plans to introduce a Code Gen Agent for Java applications. “Oracle acquired Sun Microsystems, so we would be best suited to have an agent for the Java language,” he added.

AI Quick Actions

AI Quick Actions fundamentally provide a no-code, low-code solution for deployment and fine-tuning. To help customers build, train, deploy, and manage LLMs with open source libraries such as Hugging Face’s Transformers or PyTorch, Oracle is also expanding the capabilities of OCI Data Science.

“We want to cater to a whole spectrum of data scientists and machine learning practitioners. There is a class of developers who want to try out and play with every other open source model that’s out there,” concluded Mamtani.

The post Oracle’s Generative AI Madness Begins appeared first on Analytics India Magazine.

BharatGPT on the Path to Becoming Meta of Indic LLMs

When it comes to India and generative AI, all of the recent Indic language models have been built on top of models built by Meta or Mistral. The Indian research community, though is rising up in the realm, wants to build models from scratch, and that is where BharatGPT is finding its moat.

Speaking to AIM, Ganesh Ramakrishnan, professor at IIT Bombay who is leading the BharatGPT initiative said that there is definitely a need for a foundational model for Indic languages. “We are building foundational models from scratch and that is what is keeping us busy,” said Ramakrishnan about how BharatGPT will mark India on the global AI map.

Every company and every country is racing towards being the best at generative AI, more specifically LLMs. In the beginning it was the US companies such as Google, Microsoft, and OpenAI building the best AI models, now it’s Meta with Llama 2 and Mistral from France with its open source models that can be touted as leading the way. Same goes for China with Baidu and Alibaba leading research. It is only time that India makes its mark.

Open source is the winner

The Indian academic institutions know the power of open source, especially when it comes to specialised models that are focused on healthcare or any other specific field. For example, researchers from IIT Patna recently unveiled a massive multilingual dataset for healthcare called MedSumm.

“The hope is that once we put something out, the bandwagon then begins. Folks put more stuff out and like I said, the dataset piece is very important because once you make that available, people start to utilise it in various different ways,” Aman Chadha, the researcher of the paper, told AIM.

BharatGPT is also building on with a similar philosophy. Even though BharatGPT is currently focused on building foundational models for India, the initiative has also been open-sourcing a lot of its work on Decile (similar to GitHub and Hugging Face), along with licence-permitting commercial usage, and aims to continue doing it.

Ramakrishnan also emphasised that the solutions need to be developed across different verticals as well such as banking, healthcare, farming, etc. “Mistral got France on the AI map. We want India to get on the AI map with BharatGPT.”

He said that while we can keep writing research papers and creating graduates, it is important that all these help bolster the Indian ecosystem. “I think it’s time to not just create some AI solution for India, but a working full stack,” Ramakrishnan added.

“We want everyone to use generative AI,” Vishnu Vardhan, the founder of Vizzhy, who is also the GPU buddy of BharatGPT told AIM, that the initiative would not just be about releasing weights, but making it available to everyone. He highlighted that the first model would be open source and wants developers to help them make it better. “The more people use it, the better it will become,” and eventually, they would release more versions of the model.

Vardhan emphasised that English models such as Llama and Mistral are good, but cannot address the complexity of Indian languages. He said that building Indic LLMs with a million tokens on top of English foundational models, which have trillions of token, won’t be enough to make them on par with GPT-4 for other languages such as Kannada or Tamil or Hindi.

He highlighted that Indic languages are very well connected with each other with their grammatical structure, which is very different from English. “That is why we decided that it would be easier to build models if they are originally in any Indic languages such as Hindi,” he explained, as it would be easier to mould them into other languages later.

The BharatGPT team is also working on video and speech models which would be released soon after the launch of the initial text-based foundational model.

The need for open source Indic models

Indian developers are one of the biggest contributors to open source. According to GitHub, with 13.2 million developers now on GitHub in India, the nation has firmly secured its position as the second-largest contributor to AI projects worldwide, just behind the United States. Of the total number 3.5M new developers joined GitHub in 2023.

All of this is while Meta, the open source giant is aiming for AGI goals with its upcoming Llama 3 model, and has a total of 350k H100 GPUs. BharatGPT team is also looking to acquire a lot more GPUs to continue training its models and also create an AI lab at IIT Bombay, where developers can train their own models.

“Nobody else will build these Indic language models, and it’s definitely needed at the moment,” Amit Sheth, professor at Arizona State University who recently met Prime Minister Narendra Modi for discussions on AI policy, told AIM expressing enthusiasm about the rise of Indic LLMs such as Kannada, Tamil, and Telugu models based on Meta’s Llama 2 model. But he also highlighted that India needs to build models from scratch.

Sheth also acknowledged that it is a very expensive and compute heavy task. He believes that we are not ready yet to train a model to compete with OpenAI’s GPT-4 as it would take millions of dollars and compute, which is hard to acquire at the moment. Plus, he also says that we also need to make sure about gathering data ethically.

He believes that this can be achieved with academic and private partnerships, which BharatGPT, Ola’s Krutrim, Sarvam AI, and Tech Mahindra’s Project Indus are also on the path. “A lot more research needs to be done, and I think high-quality original research is growing,” he added.

It only makes sense that BharatGPT becomes the open source foundational model for Indic language models, and as the researchers say, it is clearly headed that way.

The post BharatGPT on the Path to Becoming Meta of Indic LLMs appeared first on Analytics India Magazine.

Unlearning Copyrighted Data From a Trained LLM – Is It Possible?

Blog Image of Copyright Symbols

In the domains of artificial intelligence (AI) and machine learning (ML), large language models (LLMs) showcase both achievements and challenges. Trained on vast textual datasets, LLM models encapsulate human language and knowledge.

Yet their ability to absorb and mimic human understanding presents legal, ethical, and technological challenges. Moreover, the massive datasets powering LLMs may harbor toxic material, copyrighted texts, inaccuracies, or personal data.

Making LLMs forget selected data has become a pressing issue to ensure legal compliance and ethical responsibility.

Let's explore the concept of making LLMs unlearn copyrighted data to address a fundamental question: Is it possible?

Why is LLM Unlearning Needed?

LLMs often contain disputed data, including copyrighted data. Having such data in LLMs poses legal challenges related to private information, biased information, copyright data, and false or harmful elements.

Hence, unlearning is essential to guarantee that LLMs adhere to privacy regulations and comply with copyright laws, promoting responsible and ethical LLMs.

Stock image depicting files of copyright laws and IP rights

However, extracting copyrighted content from the vast knowledge these models have acquired is challenging. Here are some unlearning techniques that can help address this problem:

  • Data filtering: It involves systematically identifying and removing copyrighted elements, noisy or biased data, from the model's training data. However, filtering can lead to the potential loss of valuable non-copyrighted information during the filtering process.
  • Gradient methods: These methods adjust the model's parameters based on the loss function's gradient, addressing the copyrighted data issue in ML models. However, adjustments may adversely affect the model's overall performance on non-copyrighted data.
  • In-context unlearning: This technique efficiently eliminates the impact of specific training points on the model by updating its parameters without affecting unrelated knowledge. However, the method faces limitations in achieving precise unlearning, especially with large models, and its effectiveness requires further evaluation.

These techniques are resource-intensive and time-consuming, making them difficult to implement.

Case Studies

To understand the significance of LLM unlearning, these real-world cases highlight how companies are swarming with legal challenges concerning large language models (LLMs) and copyrighted data.

OpenAI Lawsuits: OpenAI, a prominent AI company, has been hit by numerous lawsuits over LLMs’ training data. These legal actions question the utilization of copyrighted material in LLM training. Also, they have triggered inquiries into the mechanisms models employ to secure permission for each copyrighted work integrated into their training process.

Sarah Silverman Lawsuit: The Sarah Silverman case involves an allegation that the ChatGPT model generated summaries of her books without authorization. This legal action underscores the important issues regarding the future of AI and copyrighted data.

Updating legal frameworks to align with technological progress ensures responsible and legal utilization of AI models. Moreover, the research community must address these challenges comprehensively to make LLMs ethical and fair.

Traditional LLM Unlearning Techniques

LLM unlearning is like separating specific ingredients from a complex recipe, ensuring that only the desired components contribute to the final dish. Traditional LLM unlearning techniques, like fine-tuning with curated data and re-training, lack straightforward mechanisms for removing copyrighted data.

Their broad-brush approach often proves inefficient and resource-intensive for the sophisticated task of selective unlearning as they require extensive retraining.

While these traditional methods can adjust the model's parameters, they struggle to precisely target copyrighted content, risking unintentional data loss and suboptimal compliance.

Consequently, the limitations of traditional techniques and robust solutions require experimentation with alternative unlearning techniques.

Novel Technique: Unlearning a Subset of Training Data

The Microsoft research paper introduces a groundbreaking technique for unlearning copyrighted data in LLMs. Focusing on the example of the Llama2-7b model and Harry Potter books, the method involves three core components to make LLM forget the world of Harry Potter. These components include:

  • Reinforced model identification: Creating a reinforced model involves fine-tuning target data (e.g., Harry Potter) to strengthen its knowledge of the content to be unlearned.
  • Replacing idiosyncratic expressions: Unique Harry Potter expressions in the target data are replaced with generic ones, facilitating a more generalized understanding.
  • Fine-tuning on alternative predictions: The baseline model undergoes fine-tuning based on these alternative predictions. Basically, it effectively deletes the original text from its memory when confronted with relevant context.

Although the Microsoft technique is in the early stage and may have limitations, it represents a promising advancement toward more powerful, ethical, and adaptable LLMs.

The Outcome of The Novel Technique

The innovative method to make LLMs forget copyrighted data presented in the Microsoft research paper is a step toward responsible and ethical models.

The novel technique involves erasing Harry Potter-related content from Meta's Llama2-7b model, known to have been trained on the “books3” dataset containing copyrighted works. Notably, the model's original responses demonstrated an intricate understanding of J.K. Rowling's universe, even with generic prompts.

However, Microsoft's proposed technique significantly transformed its responses. Here are examples of prompts showcasing the notable differences between the original Llama2-7b model and the fine-tuned version.

Fine-tuned Prompt Comparison with Baseline

Image source

This table illustrates that the fine-tuned unlearning models maintain their performance across different benchmarks (such as Hellaswag, Winogrande, piqa, boolq, and arc).

Novel technique benchmark evaluation

Image source

The evaluation method, relying on model prompts and subsequent response analysis, proves effective but may overlook more intricate, adversarial information extraction methods.

While the technique is promising, further research is required for refinement and expansion, particularly in addressing broader unlearning tasks within LLMs.

Novel Unlearning Technique Challenges

While Microsoft's unlearning technique shows promise, several AI copyright challenges and constraints exist.

Key limitations and areas for enhancement encompass:

  • Leaks of copyright information: The method may not entirely mitigate the risk of copyright information leaks, as the model might retain some knowledge of the target content during the fine-tuning process.
  • Evaluation of various datasets: To gauge effectiveness, the technique must undergo additional evaluation across diverse datasets, as the initial experiment focused solely on the Harry Potter books.
  • Scalability: Testing on larger datasets and more intricate language models is imperative to assess the technique's applicability and adaptability in real-world scenarios.

The rise in AI-related legal cases, particularly copyright lawsuits targeting LLMs, highlights the need for clear guidelines. Promising developments, like the unlearning method proposed by Microsoft, pave a path toward ethical, legal, and responsible AI.

Don't miss out on the latest news and analysis in AI and ML – visit unite.ai today.

Google Finds its Moat in Healthcare

Tech conglomerate Google is all about making an impact and seems to have a larger affinity towards life science and healthcare. Recently, Google Research launched a new medical chatbot called AMIE specialising in expert-level differential diagnosis. The chatbot creation was carried out by research leads Alan Karthikesalingam and Vivek Natarajan. Unlike the big tech’s previous AI models, Med-PaLM 2, which focuses on medical summaries or answering questions, AMIE serves as a diagnostic tool, generating differential diagnoses.

AMIE is built on Google’s PaLM and trained on datasets containing medical conclusions, summaries, and actual clinical conversations.

Last December, while OpenAI’s ChatGPT struggled to find real-life use cases, Google ventured into using LLMs for healthcare, making a tangible impact on human lives. This led to the development of Med-PaLM, an open-source LLM specifically designed for medical use.

Since then, the team released expanded versions, namely Med-PaLM-2, Med-PaLM-M, and MedLM, which are being used by different organisations to build AI-powered healthcare solutions. Currently, Med-PaLM-2 is undergoing testing at reputable healthcare institutions, including the Mayo Clinic.

Not long ago, Google DeepMind and Isomorphic Labs released the latest version of AlphaFold. The AlphaFold-latest can predict the structures from nearly all molecules in the Protein Data Bank (PDB). The AI model has already found a variety of real-life applications including finding vaccines for malaria, liver cancer, COVID-19, delivering gene therapy and more.

Pre Gen AI Boom

Even before Med-PaLM, Google, unlike its big tech counterparts, had a significant focus on democratising healthcare with technology. The company introduced a deep learning model to detect diabetic retinopathy called Automated Retinal Disease Assessment (ARDA) in 2016. The tool is still being actively employed in numerous eye clinics across India.

In an earlier interview with Forbes, Dr Karen DeSalvo, chief health officer, said, “Google is a life company…not a health company. One of the primary goals of the company is to help billions of people around the world become healthier in an equitable manner… giving every single person an opportunity to have the highest quality of health.”

Alphabet is currently focused on four key areas in healthcare, namely, wearables, health records, health-related AI, and the ambitious aim of prolonging human life.

From 2019 to 2021, a tech giant invested around $1.7 billion in healthcare projects, marking one of the highest investments by a tech company in the healthcare sector as of 2022. Since 2020, Alphabet and its affiliated venture funds have taken the lead in 25 health-related investment rounds, totaling over $1.6 billion. Overall, Alphabet and its affiliated funds have led 76 investment rounds for healthcare and biotech companies, with a combined value surpassing $4 billion.

In the last two years, the company collaborated with Mayo Clinic to better radiation therapy with AI. The company also participated in various initiatives to support pandemic response, streamline clinical workflows, and ensure secure health data management. Examples include creating a web tool with New York State for essential services access and developing a mental health wellness app with Cooper University Health Care for frontline workers.

It also partnered with Ascension to introduce Care Studio for improved patient information management, while Highmark Health integrated Verily’s digital tools for chronic condition management through an expanded Google Cloud collaboration.

Another important collaboration is between HCA Healthcare and Google Cloud, focusing on accelerating the former’s digital transformation and developing a sophisticated data analytics platform for enhanced operational models and insights, using Google Cloud’s healthcare data resources.

Leading the AI in Healthcare Race

It is not like other big techs like Microsoft, AWS, Apple are not tapping into healthcare. However, it was Google who started the “AI and healthcare” conversation last year.

Since then, all the big techs have decided to take this area seriously. For example, AWS has collaborated with prominent healthcare and pharmaceutical companies, such as Pfizer, Amgen, Moderna to enhance drug discovery, manufacturing, and AI integration. Apple also plans to introduce additional health detection features in its next series of watches, addressing issues like hypertension and apnea.

Oracle is also adding generative AI solutions in its cloud-based electronic health record (EHR) capabilities that include enhancing patient care, optimising clinical expertise, increasing cost efficiency, and reducing provider burnout.

Moreover, OpenAI also got into this growing space by recently partnering with wearable technology company WHOOP to create a personalised health and fitness coach driven by its LLM GPT-4.

However, another interesting this to note here is that most major tech companies depend on partners for research and subsequently apply findings to projects, Google does both in-house. No other conglomerate appears to have invested as extensively in developing LLMs specifically for healthcare.

According to Vivek Natarajan, AI researcher at Google and one of the main brains behind AMIE, the tech giant’s goal is to be helpful to people and health is a big part of people’s lives.

“AI is going to be a big part of healthcare; with Google’s expertise in AI, there is no better place for medical AI research. At the same time, we also deeply care about going beyond research to translational and having an impact on people’s lives,” Natarajan told AIM.

Read more: Meet the Genius behind Med-PaLM 2

The post Google Finds its Moat in Healthcare appeared first on Analytics India Magazine.

Linguist Emily M. Bender Has a Word or Two About AI

“Who are they to be speaking for all of humanity?,” asked Emily M. Bender, raising the question to the tech companies in a conversation with AIM. “The handful of very wealthy (even by American standards) tech bros are not in a position to understand the needs of humanity at large,” she bluntly argued.

The vocal, straightforward, and candid computational linguist is not exaggerating as she calls out the likes of OpenAI. Currently, Sam Altman is trying to solve issues of humanity, which include poverty, hunger, and climate catastrophes through AI tools like ChatGPT, which has been developed in Kenyan sweatshops, got sued for violating privacy laws, continues to pollute the internet and is a source of misinformation.

“I would love to see OpenAI take accountability for everything that ChatGPT says because they’re the ones putting it out there,” she said without hesitation, even though it has been long debated who should bear the blame – developers or users, when technologies backfire.

She sternly adds that they are the ones who set up the means to spill synthetic information into the ecosystem. So far, there’s no accountability for that, and there should be,” demanded the 50-year-old professor who resists being called an AI researcher. Why? “Because I’m not interested in the project of building AI. But I collaborate with many people who fit under the umbrella of AI ethics,” the industry critique clarified.

Bender has been long busting the over promises of AI, but she believes in the value of language technology. (Even though she has never used ChatGPT!)

“There’s a lot of speakers and people using [language technology]. When designing technology, we need to think about who we are designing it for,” she stated. If someone has built a text or speech system for American English, they should not be trying to sell that system in the Indian market. That doesn’t make sense. She thinks machine translation can be handy when there’s transparency about what it’s good at and not good at.

Bender pointed out that one of the issues with machine translation is that if the output sounds fluent, the user is more likely to think that it’s also accurate. “Those two things don’t necessarily correlate,” she highlighted.

Language technology can be beneficial in search engines. We still need search engines where you can prompt in any natural language. “I would like to find bread recipes without salt and have it understand what the without means and respect that in the query because that’s not how search engines are built,” Bender (who has written about the subject) frowned.

There’s a lot of work to be done here and a lot of value. But there’s all this energy and focus on large language models, which were labelled stochastic parrots in a paper Bender co-authored. They called the technology so because it spits out plausible-sounding text with no basis in any communicative intent.

Regulating Actions

“You can have all the good intentions in the world, but you’re not going to get very far until there’s some regulation that protects the rights that the profit motive runs roughshod over,” Bender dropped another truth bomb.

“These are rights of people as individuals and also as communities,” she continued and suggested having better regulations, and regulated dialogues around ethics.

The companies selling AI have a meaningful seat at the table of law. Hence, Bender has been talking to policymakers from the United Nations and its financial arm, the International Monetary Fund. Over time, she has discovered that the problem is that regulators are also hearing from the technologists, and the lobbyists are paid to “make sure the regulation doesn’t get in the way.”

In the US context, she often hears regulators saying they have to make sure not to hinder innovation. After that, they will consider people’s rights. Most of Bender’s work reminds them [the regulators] that they’re there to protect the rights of the people, not the rights of the corporations.

People who question the authorities don’t want to inhibit innovation, she clarified. “We want to push all of this energy and capital spent on AI models in the directions that benefit people and not trample their rights,” she explained, having worked with superstar ethicist Timnit Gebru, whom Google fired for pointing out the problems of in-house AI models.

“Regulation can channel innovation in that way,” suggests Bender. She continues to explain that the penalties by regulators must be big enough to dissuade the companies from malpractices.

When she was a kid, one of the environmental issues was the hole in the ozone layer. “It’s not a problem anymore because we created regulations,” she explained. Probably, innovation came up with other solutions, and it is not a problem anymore, she added.

The second hurdle is that it is difficult to have conversations about ethics because people ask, “Why are you telling me what to do?” One way to circumvent that kind of objection is by acknowledging their good intentions and discussing the possible impacts of this technology, Bender suggested.

Reminiscing the AI-Free Google

Bender recalled that Cory Doctorow coined the term ‘enshitiffication‘, where “you have a decent platform, and then its owners make it worse for everybody else involved. We’re (definitely) seeing that with Google search results.”

In the original Google paper about PageRank, they said that if you add an advertising incentive to this, it will not work. Then it became an advertising company,” Bender sighed.

“Back when it started, it was an effective way to find information on the web. Then, between their advertising incentives and search engine optimization, it got harder and harder to use. The condition is worsening further, with Google and others putting out synthetic stuff all over the place. It’s getting even harder to find good information,” Bender said.

She is inspired by the work of Sophia Noble, the author of Algorithms of Oppression, who asks us to think about information access as a public good rather than something that should be in private interests.

“I don’t think we have to believe the tech companies that just because this is the path we’ve been on, it’s the only path,” Bender rightly concluded.

The post Linguist Emily M. Bender Has a Word or Two About AI appeared first on Analytics India Magazine.

Why Universities Need OpenAI 

Recently, OpenAI partnered with Arizona State University (ASU) to integrate ChatGPT into its classrooms. The main aim of the university is to enhance student success, foster innovative research, and streamline organisational processes.

The college said that the use of generative AI on campus will be guided by the faculty members and the ChatGPT usage will begin in February for the project submissions. ASU said that some professors already use AI in composition and journalism classes, and chatbots may even serve as personalised tutors.

The university further plans to use ChatGPT Enterprise to develop AI avatars as a “creative buddy” for studying certain subjects, like bots that can sing or write poetry about biology, for instance.

Interestingly, this is the first instance of a university partnering with OpenAI. When ChatGPT was launched, many universities across the globe banned it, citing concerns that students were using it to copy their assignments. Professors implemented various AI detection tools to identify plagiarised submissions, however, it didn’t turn out to be a foolproof solution as the detectors weren’t accurate.

ChatGPT Enterprise for Education

In the latest episode of ‘Unconfuse Me with Bill Gates,‘ OpenAI Chief Sam Altman pointed out that healthcare and education are among the key areas he is excited about using ChatGPT.

A recent survey found that 30% of college students used ChatGPT for schoolwork during the 2022-23 academic year. While the specific number of universities using ChatGPT is not provided, it is clear that the technology has made a substantial impact.

The move by Arizona State University to use ChatGPT Enterprise might act as a catalyst for others to follow suit. Although the concept of an AI tutor is not new, ChatGPT Enterprise has made the creation process much more accessible. OpenAI’s recent launch of the GPT Store allows users to craft personalised GPTs using their own data.

Considering the abundance of data within universities, including notes, lectures, presentations, and quizzes, professors can harness this information to develop customised GPTs for specific subjects. This, in turn, assists students in resolving their doubts and queries effectively.

ChatGPT Enterprise will provide students as well as educators with a number of advantages. It will offer aid with research, writing, tutoring, learning, ideation, brainstorming, language translation, programming, virtual office hours, accessibility support, professional development, and fast responses to frequently asked questions.

A user on X wrote, “Digital professors can factually explain subjects without emotion or most cognitive biases, giving students room to form their own opinions,” saying that virtual instructors won’t harass, intimidate, or judge students. Moreover, AI mentors can spend hours with every student to help with homework, research, creative inspiration, writer’s block, time management, motivation, budgeting, etc.

In addition to these advantages, ChatGPT Enterprise may assist with language translation, facilitating communication and comprehension between many languages.

It can also aid students studying computer science or programming by addressing coding issues, troubleshooting, and helping them comprehend difficult ideas. For example, Khan Academy’s Code Tutor helps students expand their coding skills.

ChatGPT Enterprise can further assist students in drafting their college application essays. For instance, a student recently created a GPT called University Application Advisor that helps with university applications. However, it doesn’t actually write essays but only provides feedback.

OpenAI took a step last year to address the overlooked teacher segment. The initiative involved the announcement of a new guide, enabling teachers to effectively utilise ChatGPT in classrooms. OpenAI provided prompts and various examples to guide teachers on making the best use of ChatGPT in teaching students.

Dr. Helen Crompton, Professor of Instructional Technology at Old Dominion University, encourages her education graduate students to use ChatGPT as a stand-in for a particular persona—like a debate partner who will point out weaknesses in their arguments, a recruiter who’s interviewing them for a job, or a new boss who might deliver feedback in a specific way.

OpenAI’s partnership with Arizona State University is just the beginning. If other universities come out with similar announcements, it wouldn’t be a surprise, considering the benefits of AI in education.

The post Why Universities Need OpenAI appeared first on Analytics India Magazine.

Tamil Llama Creator Unveils Malayalam and Telugu Llamas

In a recent update, Abhinand, the creator of the Tamil LLaMA, has included support for Telugu and Malayalam, enhancing its performance over Meta’s LLaMA 2 across various benchmarks. This development builds on the success of the original Tamil LLaMA project, released on GitHub in November 2023.

To address limitations and widen the scope, Abhinand introduced Tamil LLaMA v0.2. This bilingual model excels in English and Tamil, marking a significant step forward.

The team also had JarvisLabs.ai’s support with GPUs, facilitating the development of Telugu and Malayalam LLaMA models. These models are now accessible on HuggingFace Hub.

The model adaptation process involved pretraining, fine-tuning, and alignment. The pretraining phase focused on expanding vocabulary and enhancing language generation capabilities. Fine-tuning involved training on a substantial set of instruction-response pairs, while alignment ensured human-preferred responses using techniques like RLHF and DPO.

Despite the alignment stage, the models remain largely uncensored.

Inspired by Sarvam AI’s OpenHathi, Abhinand’s approach took a distinct technical route. Improvements were made to the Tamil LLaMA tokeniser, and a comparative analysis against Indic Language LLMs on English benchmark scores was conducted. Based on the Open LLM Leaderboard, the evaluation positioned the new models favourably against LLaMA 2 and OpenHathi.

The fine-tuning stage aimed to match or surpass the original LLaMA 2 model’s English performance while enhancing language abilities in Tamil, Telugu, and Malayalam. This involved fine-tuning on a vast corpus of instructions, creating a synthetic dataset for regional knowledge, and performing DPO for further enhancement.

The result is the Tamil LLaMA v0.2 model marginally outperforming LLaMA 2 Chat on various benchmarks, showcasing advancements in linguistic capabilities.

The Tamil LLaMA project, initiated in September 2023, aimed to adapt the features of LLaMA 2 for the Tamil language. The project’s success led to the release of 7B and 13B parameter model variants. Open-sourcing the project facilitated collaboration and adaptations for other Indian languages like Hindi, Odia, and Kannada.

In an exclusive interview with AIM, Balachandran revealed the genesis of the Tamil LLaMA project, citing inspiration drawn from the Chinese LLaMA Alpaca model.

He stated, “Chinese is a bit of a complex language, but if they can make it work for Chinese, then surely we will also be able to make it work for Indian languages, right? So that was the motivation,” emphasised Balachandran.

The post Tamil Llama Creator Unveils Malayalam and Telugu Llamas appeared first on Analytics India Magazine.