The branches of Machine Learning and AI
In the past few years, artificial intelligence (AI) has emerged as a formatrice force across research and industry. This field not only redefines how machine and data-driven algorithms can perceive, exploit and transform information. In this article, we embark on a short journey into the main types of learning approach techniques: supervised learning, self-supervised learning and reinforcement learning.
Supervised learning:
Supervised learning is similar to a guiding hand for machines. The model is trained using label data meaning that it is provided input/output pairs and learns the mapping function between those pairs. This learning approach is most prevalent when a significant amount of pre-labeled data is available in tasks like object recognition, where a model learns to identify objects based on pre-labeled images.
The success of supervised learning is closely related to the quality and diversity of the training data, as it directly influences the model’s ability to generalise to unseen data. Descript its effectiveness, supervised learning faces challenges in situations when only a limited amount of data is labelled or when the cost of labelling is too high. To circumvent these limitations, some research papers have shown that one can leverage a pre-trained model and use it as a general feature extractor or semi-supervised techniques.
Self-supervised learning: “The dark matter of intelligence”
Yan LeCun interview with Lex Friedman: https://www.youtube.com/watch?v=SGzMElJ11Cc
Majors reference papers:
- Bootstrap Your Own Latent https://arxiv.org/abs/2006.07733
- Barlow Twins https://arxiv.org/abs/2103.03230
Unlike the structured guidance framework of supervised learning, unsupervised learning can be applied to unlabeled data. The task of the learning algorithm is to identify and uncover patterns, structures or relationships without explicit instructions.
Clustering and dimensionality reduction and common applications of unsupervised learning. In clustering, the algorithm groups similar data points together, revealing an inherent structures while dimensionality reduction can be viewed as a compression technique that aims to simplify complex data while preserving essential information, aiding in visualisation and analysis.
One significant challenge in unsupervised learning is the absence of ground-truth meaning to evaluate the effectiveness of the model. The interpretability of results becomes crucial and researchers explore visualisation techniques like TNSE and metrics like Davies-Bouldin index to assess the quality of clustering.
Reinforcement Learning:
Majors agents:
- AlphaGo: https://deepmind.google/technologies/alphago/
- OpenAI gym https://openai.com/research/openai-gym-beta
Reinforcement learning (RL) mirrors how living beings learn through trial and error. An algorithm known as an agent interacts with an environment and learns to make decisions by receiving feedback for any action taken in the form and reward or penalties. By accumulating the feedback from multiple actions, the agent refines its strategy, known as policy, in order to maximise its total rewards. Many parallels can be drawn between and pre-existing fields like Optimal Control and RL but key differences lie in the absence of a well-defined model of how the world evolves through time.
Games like Go and Atari have shown the capabilities of RL. The agent learns the optimal policy by exploring different actions and observing their consequences. The balance between exploration and exploitation is a key challenge in RL, as a fully explorative agent may struggle to converge to a successful policy and a fully exploitative agent may not find the best policy.
Ongoing researches in RL put their focus on enhancing the sample efficiency to address long-term memory and extending its applicability into real-worlds domains.
Conclusion:
Machine Learning, with its diverse forms, has become a cornerstone of advancements in AI. From the structured guidance of supervised learning to open exploration of self-supervised learning and the dynamic adaptability of reinforcement learning, each type brings unique strengths and challenges. As research pushes boundaries of these frameworks, ML continues to pave the way for intelligent systems able to learn from data and show adaptability to unseen scenarios.
