mian ahtisham’s Post

🎓 Completed Machine Learning Course I’m excited to share that I’ve recently completed a comprehensive course in machine learning. The course covered key concepts including algorithms, data analysis, and model building, equipping me with valuable skills to leverage AI in various applications.

Came across an incredible resource that brings machine learning concepts to life through stunning visuals! 🎉 To all my #DevOps, #CloudEngineers, #SystemAdmins, if you've ever felt overwhelmed by dense theory, these interactive essays are engaging like fresh air. Dive into neural networks, decision trees, and more in a way that's engaging and accessible. https://lnkd.in/gHhCkhXp 🚀 Check it out and let's explore together. #MachineLearning #AI #MLU #DataScience #TechEducation #AIExplained #GitHub

Contrary to popular belief, claims of explainable/interpretable AI do not mean those explanations are meaningful to users. -- "What we have essentially found is that there is a subdiscipline of computer science that has been enthusiastically applying their methods to the development of machine learning models under a largely unsupported assumption that their models enabled better human interpretability.” Explainability in AI is not a binary "it has it or it doesn't" feature. This is akin to whether a car comes with airbags or not. Sure, airbags are a required safety feature, but not all airbags are created equal. If we treated airbags like our AI explainability claims, then some cars may as well just have party balloons in the steering wheel. Is the feature there? Sure. Does it provide a meaningful impact? Nope. Just because an AI model is claimed to be explainable, does not mean that it's explainable in a meaningful way. We must not take claims of AI explainability at face value. Just as AI performance must be rigorously proven, claims of AI explainability need to be demonstrated as well. Keep doing good work Hosea Siu, Kevin Leahy, Makai M. #AI #explainableAI #machinelearning #responsibleAI

Our machine learning interpretability work has been published as a case study for broad audiences by the MIT Social and Ethical Responsibilities of Computing initiative. Key insight: "It is standard practice in machine learning to present performance metrics for new models and algorithms, typically with some values of model accuracy... However, it is still very common in machine learning literature to call an approach “interpretable” or "explainable” or “transparent” without defining what is meant by those terms and without evidence that the definitions are met. Nonetheless, when models are presented (or sold) to end users—everyone from hospitals to factories to militaries — these properties are typically touted as important differentiators for why one model is better than another." Or, more bluntly - in ML world, call a model "accurate" without evidence and you get laughed out of the room. Call a model "interpretable" without evidence and you get invited to a conference!

The right inductive biases help increase the performance of ML algorithms. But you can also use inductive biases to proactively improve model interpretability, robustness, and plausibility. Read more on Mindful Modeler:

In the field of causal inference (CI) and machine learning (ML), we can distinguish between: 1. Causal inference for machine learning and, 2. Machine learning for causal inference Example of ML for causal inference: meta-learners (s-learner, t-learner, x-learner) etc. Example of causal inference for ML: causal discovery, DAGs, etc. ML for Causal Inference: Applying ML techniques to estimate causal effects, identify causal relationships, and make counterfactual predictions. This is about using the power of ML to solve problems traditionally addressed by causal inference methods. Causal Inference for ML: Integrating causal reasoning into the ML workflow to enhance model interpretability, fairness, robustness, and generalization. This is about improving ML models by leveraging insights from causal inference.

Importance of Probabilistic Models in Machine Learning

It may be easier than you think to use your skills for server-based deep learning on Apple devices. Yunfei Cheng and I attempted to evaluate the learning curve by comparing MLX kernels working on Metal GPUs in Apple Silicon chips to PyTorch kernels on CUDA GPUs. The image below depicts the scalability of self-attention and linear projection on M1 Max, M2 Ultra, A100, and H100. The x-y plane represents the beam shape size used in our Recurrent Drafting work (https://lnkd.in/dvrvUwbU). All of these kernels show a similar scalability trend as the beam shape grows. It is interesting to reveal that the performance difference between CUDA and Metal in SDPA is considerably lesser than in linear projection. For example, linear projection indicated a roughly 100x performance difference between the M1 Max and the H100, whereas SDPA showed just a 25x difference on the same hardware.

Mastering Hyperparameter Tuning for Optimized Machine Learning Models Hyperparameter tuning is the secret sauce that transforms a good machine learning model into a great one. By fine-tuning parameters like learning rate, tree depth, or number of layers, you can maximize performance and accuracy. Key Highlights from the Article: 1. What is Hyperparameter Tuning? A method to optimize non-learnable parameters in a machine learning model. Impacts training speed, convergence, and overall accuracy. 2. Techniques for Tuning: Grid Search: Systematic exploration of parameter combinations. Random Search: Random sampling of hyperparameters for efficiency. Bayesian Optimization: Intelligent exploration for fewer iterations. 3. Practical Steps with Code: Learn how to implement tuning using libraries like Scikit-learn, TensorFlow, or PyTorch. Understand real-world examples of hyperparameter tuning in action. 4. Challenges: Time-consuming process for large datasets. Risk of overfitting when tuning excessively. https://lnkd.in/gkBMQ4vc Additional Resources: Tools for Automation: Optuna, Ray Tune https://lnkd.in/gExMuTnF Code Examples: Explore hyperparameter optimization on GitHub. https://lnkd.in/g3n_SQbp #MachineLearning #HyperparameterTuning #AI #DataScience #MLModels #OptimizationTips

Qwen2-Math Released: A Comprehensive AI Suite Featuring Models Ranging from 1.5B to 72B Parameters, Transforming Mathematical Computation The Qwen Team has recently released the Qwen 2-Math series. This release, encompassing several model variants tailored for distinct applications, demonstrates the team’s commitment to enhancing AI’s proficiency in handling complex mathematical tasks. The Qwen 2-Math series is a comprehensive set of models, each designed to cater to different computational needs. The lineup includes: ✅ Qwen 2-Math-72B ✅ Qwen 2-Math-72B-Instruct ✅ Qwen 2-Math-7B ✅ Qwen 2-Math-7B-Instruct ✅ Qwen 2-Math-1.5B ✅ Qwen 2-Math-1.5B-Instruct These models vary in complexity and instruction-following capabilities. It provides users with various options depending on their specific requirements. At the top of the range is the Qwen 2-Math-72B, a model that boasts an impressive 72 billion parameters. This variant is designed for highly complex mathematical computations and is suitable for tasks requiring deep learning and extensive data processing. The “Instruct” version of this model, Qwen 2-Math-72B-Instruct, offers additional enhancements that allow it to follow user instructions more precisely...... Read our full take on Qwen 2-Math: https://lnkd.in/gD5EjZzB Check out the models here: https://lnkd.in/gdZ_H52Q Alibaba Group Alibaba.com