AI Techniques for the Modern Problem Solver
A curated list of AI tools and techniques to solve problems. No background knowledge assumed, ideal for newcomers to the field.
Here AI is an umbrella term for the ability to adaptively and efficiently achieve goals via artificial learning, planning and decision making, including techniques from Statistics, Data Science, Machine Learning, Artificial General Intelligence and related fields.
The goal is to let you know that these techniques exist, giving a basic usage example, resources and weaknesses.
Contributions to the list are welcome.
Template
Superpower: What it does.
How hard: How hard to use it without understanding the theory behind (debugging, parameter selection, etc.)
Libraries: Suggested implementations.
Example: Basic tutorial.
Weak Points: When it tends to not work well.
Other: Other!
Table of Contents
- Neural Network Classifier (Supervised Learning, Deep Learning)
- Neural Network Object Detector (Supervised Learning, Deep Learning)
- Semantic, Instance and Panoptic Segmentation (Supervised Learning, Deep Learning)
- Transfer Learning (Deep Learning)
- Decision Trees (Supervised Learning)
- Ensemble methods
- Random Forest (Supervised Learning)
- XGBoost (Supervised Learning)
- AutoML (Supervised or Unsupervised Learning)
- Clustering (Unsupervised Learning)
- Data Augmentation
- Bayesian Inference and Probabilistic Programming
- Distribution Fitting
- Anomaly Detection
- Graphs
- Graph Neural Networks
- Null-Hypothesis Significance Testing
- Reinforcement Learning
- Genetic Algorithms
- Time Series Forecasting (Classic)
- Online Learning (a.k.a. Incremental Learning)
- Monte Carlo Tree Search
- Extra
Superpower: Classify objects in images, text, video and more.
How hard: Medium. Popular defaults are good enough to get far, the bottleneck is getting good data. Debugging is tricky.
Libraries: Detectron2.
Example: Vision classifier on CIFAR10 dataset
Weak Points: Needs a lot of labelled data to be highly accurate, hard to interpret, hard to quantify prediction confidence, long training time.
Other: Top choice to classify high level concepts such as "cat vs dog" given an image.
Superpower: Detect objects in images and videos, usually returning a bounding box.
How hard: Same as Neural Network Classifier.
Libraries: Detectron2, yolov8.
Example: Object Detection with Detectron2
Weak Points: Same as Neural Network Classifier.
Other: Production ready tool, working in real time.
Superpower: Detect instances of objects inside an image, pixel per pixel. Semantic segmentation treats them as aggregates detections, while Instance segmentation singles out individual objects but do not consider every single pixel. Panoptic segmentation combines both.
How hard: Same as Neural Network Classifier.
Libraries: Detectron2.
Example: Instance Segmentation with Detectron2
Weak Points: Data labelling is very time consuming.
Other:
Superpower: Train a neural network with little data, by reusing a pre-trained network as a starting point.
How hard: Easy.
Libraries: Your favorite DL library would do.
Example: Transfer Learning For Computer Vision
Weak Points: May inherit some restrictions from the original model.
Other: New predicted classes should be somehow similar to the pre-trained ones for better results. Otherwise just train longer.
Superpower: Classification and regression with little data preparation and no (or few) parameters.
How hard: Easy.
Libraries: Sklearn.
Example: Decision Trees with Sklearn
Weak Points: Tend to overfit or be unstable, not suited for high-dim data (e.g. images).
Other: Great explainability (can be mapped to yes/no questions), great for tabular data.
Superpower: Combine multiple models to reduce overfitting.
How hard: Easy.
Libraries: sklearn.ensemble
Example: Single estimator versus bagging: bias-variance decomposition
Weak Points: Algorithm complexity and interpretability are sacrificed for better accuracy, but the improvements are usually not that large.
Other: Used to squeeze few additional accuracy points and win machine learning competition, e.g. on Kaggle. Important examples are Random Forests and Gradient XGBoost.
Superpower: Classification and regression with little data preparation for tabular data.
How hard: Easy.
Libraries: Sklearn.
Example: Random Forest with Sklearn
Weak Points: Less explainable than decision trees, not suited for high-dim data (e.g. images).
Other: Less overfit than decision trees, great for tabular data.
Superpower: Classification and regression with little data preparation for tabular data.
How hard: Easy.
Libraries: XGBoost or using Sklearn.
Example: XGBoost in Python
Weak Points: Less explainable than decision trees, not suited for high-dim data (e.g. images).
Other: Less bias and variance than decision trees, more scalable. Alternative ensembling strategy to random forests, using Gradient Boosting instead of Bagging. Consistently at the top of machine learning competitions.
Superpower: Automate the algorithm selection, remove human bias.
How hard: Easy.
Libraries: auto-sklearn
Example: Minimal AutoML tutorial
Weak Points: Very slow, lack of control.
Other: Can make sense for the initial exploration of problems with no clear angle of attack.
Superpower: Unsupervised clustering of data into classes.
How hard: Easy.
Libraries: sklearn.cluster
Example: K-Means clustering on the handwritten digits data
Weak Points: Not as good as deep learning for complex problems.
Other: Many flavours of clustering are available.
Superpower: Increase the size of the dataset.
How hard: Easy.
Libraries: Pytorch.
Example: Data Augmentation with Pytorch plus many format specific. For instance Albumentations
Weak Points: Not really, it is always suggested to perform some data augmentation.
Other: Often the best data augmentation is specific to your problem. Think at how can you generate realistic synthetic data programmatically.
Superpower: Lean formalism to perform inference given prior knowledge, backed by solid theoretical understanding.
How hard: Medium, mainly since you need to get familiar with some statistics jargon.
Example: PyMC Basic Tutorial
Weak Points: As always in statistics, the reliability of the model is dependent on the underlying assumptions made, for instance on the priors.
Other: Popular algorithms include Markov chain Monte Carlo.
Superpower: Understand if your data is well described by a known distribution.
How hard: Medium, some knowledge of probability distributions is needed.
Libraries: Scipy or wrappers like FITTER
Example: Scipy curve fitting
Weak Points: The fitted function has always less information that the data itself.
Other: Popular fitting distribution are linear, Gaussian and Poisson.
Superpower: Detect outlier data-points.
How hard: Easy.
Example: Comparing anomaly detection algorithms for outlier detection on toy datasets.
Weak Points: To avoid false alarms the base dataset needs to be clean, large and comprehensive. The latter is particularly challenging, as many problems suffer from distribution imbalance.
Other: The alarm sensitivity is problem specific. How costly is a false alarm for your use case?
Superpower: Map the problem to a graph, to exploit fast and battle-tested algorithms available for graphs.
How hard: Easy if not using deep-learning, medium otherwise.
Libraries: NetworkX
Example: NetworkX Basic Tutorial
Weak Points: Modelling the system with a graph may require unrealistic simplifications.
Other: Graphs have been widely studied in mathematics and computer science, many optimized algorithms exists.
Superpower: Use deep learning on graph-like data natively, without loss of information.
How hard: Hard, due to the relative novelty of the field.
Libraries: Pytorch Geometric
Example: Hands-on Graph Neural Networks with Pytorch Geometric
Weak Points: Some graph structures cannot be distinguished by GNNs even in principle.
Other: GNNs are a special case of Geometric Deep Learning, that is deep learning on non-euclidean data (graphs, manifolds).
Superpower: Confirm or reject a hypothesis.
How hard: Medium, p-values can be confusing and are sometime used maliciously (p-hacking).
Libraries: Not needed.
Example: Testing the fairness of coin
Weak Points: Reporting the whole statistics, or at least multiple estimators, is more accurate than just the p-value.
Other: Commonly used in scientific literature.
Superpower: Learning how to act optimally in a dynamic environment in order to maximize a reward.
How hard: Hard. Hard debugging. It may be hard to create a meaningful reward function. Even top algorithms can be unstable.
Libraries: Stable Baselines 3. RLib is more production-ready, but also less user friendly. Check also the Gym Environments docs.
Example: Cart Pole Environment in Stable Baselines 3
Weak Points: Slow to train, needs a lot of data. Requires a good simulator of the environment.
Other: Applicable to vastly different problems, from nuclear fusion to pure maths. State of the art techniques are usually deep learning based (Deep Reinforcement Learning).
Superpower: Optimize parametrized functions by biologically-inspired mutation, crossover and selection of candidate solutions.
How hard: Hard to write an accurate and scalable fitness function.
Libraries: PyGAD.
Example: Genetic Optimisation in PyGAD
Weak Points: Compute intensive by design, usually needs an approximate fitness function. Do not scales well with complexity.
Other: Used for hyperparameter optimization.
Superpower: Predict future values of a time-series based on past results.
How hard: Easy.
Libraries: SkTime.
Example: Forecasting with sktime
Weak Points: Not able to predict over long periods due the simplifications introduced in the model. An alternative to building increasingly complex models is use deep learning based models.
Other: Favoured over neural network based models when explainability is important. ARIMA is the most popular class of forecast models.
Superpower: Learn from a stream of data, without the need for retraining the whole model.
How hard: Medium.
Libraries: River.
Example: From batch to online learning
Weak Points: Often batch training is just fine and more pragmatic. Online is less accurate than batch learning if the data has no concept drift.
Other: Great to incorporate data from recent events into your model, for instance for LLMs. But it is still hard to do it well with neural networks.
Superpower: Search the best option by simulating outcomes, without requiring heuristics.
How hard: Hard.
Libraries: LightZero.
Example: MuZero agent to play CartPole
Weak Points: Compute intensive. May miss pivotal moves leading to suboptimal strategies. Pretty complex to setup, usually paired with neural networks to perform predictions, RL to improve from self-play.
Other: Used to solve many two-players games, such as Chess and Go.
Other relevant tools or techniques include (may be included in the above in the future):
- Kalman Filters
- Monte Carlo Estimations
- Support Vector Machines
- Alpha-Beta Pruning
- Logistic Regression
- Dimensionality Reduction
- Generative Models (GANs, VAE, Diffusion)
- Sim to Real
- Sequence Models (LSTMs, RNNs, Transformers)
- Multi agents simulations
- Knowledge Representations
- Knowledge Distillation
- Domain adaptation
- Reinforcement Learning from Human Feedback (RLHF)