The model shelf
AI
Models I have built, trained or put into production, and the ideas under them. Written for people who like seeing how it works.
Things I run and train
- SpiRobs reach policyTraining A reinforcement-learning agent that steers a three-tendon arm to a target Simulator MuJoCo, 3DTask Reach, tip to a target pointActuation 3 tendon actuators
- Self-improving voice promptsIn production Thompson Sampling over four live prompt variants, scored on five rubrics Selection Thompson Sampling, BayesianVariants 4 competing prompts on live trafficScoring 5 rubrics, clarity to tone
Rebuilt from scratch
Reimplemented without a framework, to see how it works.
- GPT-2 from scratchBuilt from scratch A mini transformer trained on Tiny Shakespeare Framework PyTorchParts Tokenizer, embeddings, causal multi-head attention, feed-forwardData Tiny Shakespeare, character level
- BPE tokenizerBuilt from scratch The cl100k_base tokenizer, reimplemented Language PythonTarget cl100k_base, as used by GPT-4Covers Merge rules, vocabulary construction
- Word2VecBuilt from scratch Word embeddings from scratch, bag-of-words, 5000 words in 60 dimensions Approach Bag-of-wordsVocabulary about 5000 wordsEmbedding size 60 dimensions
- BrainCellsBuilt from scratch A neural network from raw PyTorch tensors: no torch.nn, no autograd Framework PyTorch tensors onlyWritten by hand Forward pass, activations, loss, backprop, weight updatesCompanion Detailed blog on neural networks
- ML from scratchBuilt from scratch Perceptron, softmax regression, GDA and SGD with only NumPy Stack NumPy and raw mathImplemented Perceptron, softmax regression, Gaussian discriminant analysis, SGDStatus Ongoing, still adding
- Naive Bayes spam classifierBuilt from scratch Classifying spam from word frequencies, with no ML library Stack NumPy, SciPy, PandasTask Spam or not spam from word frequenciesCore idea Bayes theorem with independent words
- Edge object detectionBuilt from scratch On-device vision driving an ESP32 Hardware Jetson Nano and ESP32Loop Detect, decide, move
Code on GitHub
The repositories behind all of this. All of GitHub ↗
- SpirobsRoboticsA tentacle built with actuators and cables for tendonsPython
- NLPAIAll my NLP projects and learnings, including the GPT-2 buildJupyter Notebook
- Word2VecAIWord2Vec from scratchJupyter Notebook
- BrainCellsAINeural networks implemented from scratchPython
- from_scratch_implementationsAIBasic ML concepts implemented from scratchJupyter Notebook
- Research-PapersAINotebooks and codeJupyter Notebook
- Object-detectionAIAn object detection model that identifies signs made on the webcam