Machine Learning and Deep Learning: A Map of Signals, Models, and Tests

Re-edited on July 30, 2026, this article replaces a fictional ladder with a map: deep learning is a family within machine learning, not a parallel field or a step that leads by itself to artificial general intelligence. Supervised, self-supervised, and reinforcement learning describe where the signal for adjusting a system comes from; “deep” describes the type of model.

The right choice starts with the problem rather than a fashionable technique. A team must state which input exists, which output is needed, who provides feedback, which error matters, and how the environment will change. A model can learn from millions of examples and remain specific; it can also require extensive human work even when labels are not written manually.

A map with axes, not a hierarchy

In supervised learning, each example contains an input and target output: image and diagnosis, message and category, house and price. An algorithm adjusts a function to reduce error on the training set. The test uses separate data and should represent the future population. If information from the target leaks into features or the same people occur in different splits, the score is inflated.

Unsupervised learning seeks structure without an equivalent target label, such as groups, densities, or factors. Self-supervised learning constructs a task from the data itself, perhaps hiding one part and predicting it or comparing two transformations of an example. Reinforcement learning uses consequences of actions over time. These regimes can be combined: a model may be pretrained through self-supervision, tuned with labels, and then used to learn a policy from feedback.

LeCun, Bengio, and Hinton’s Deep learning review describes models with multiple representation levels that transform raw data into progressively abstract features. “Deep” refers to that composition of layers. It does not mean an explanation is deep, that a system mimics brain connectivity, or that its knowledge is general.

Three questions that commercial language often blends should be kept apart. Architecture describes which transformations a model can represent; the learning regime describes which signal changes its parameters; and the complete system describes how data, tools, and decisions are assembled. The same network can be trained with labels, through a self-supervised task, or inside a reinforcement-learning agent. An application may also combine a network with explicit rules, search, and human controls. Calling it “deep learning” does not identify any of those choices by itself.

Supervision does not remove human work; it moves it

Labels come from experts, users, sensors, records, or rules. Every source contains decisions and errors. Before training, the unit, period, classes, and treatment of disagreement must be defined. A clinical label may depend on a later test; a fraud category may emerge months afterward; a user action may reflect screen position rather than preference.

Deep networks can learn features and a predictor jointly, reducing manual engineering in some problems. They still embody human choices about data, loss, architecture, and threshold. Representation Learning reviews criteria and methods for learning representations, including the usefulness of explanatory and transferable factors. A learned representation is not guaranteed to match causes in the world or retain what another task needs.

A minimum control splits training, validation, and testing by the unit that will be new in production—person, device, location, or time—and preserves a simple baseline. Linear regression, trees, or rules may outperform a network when data is limited, tabular, or stable. Complexity is justified only when it improves the relevant metric after latency, maintenance, and reviewability are included.

Self-supervision: manufacturing a signal is not assumption-free learning

BERT masked tokens and trained a model to recover them before fine-tuning it on labeled tasks. SimCLR learned visual representations by drawing two transformations of the same image together and pushing others apart. In both cases, designers chose what to hide, which transformations preserve identity, and which examples count as comparisons.

An autoencoder learns to reconstruct an input after passing through a representation. It can compress or remove noise when its bottleneck and loss encourage that behavior. It can also learn shortcuts and preserve irrelevant detail. Good reconstruction does not establish that the code represents disease, intent, or meaning; usefulness must be tested on later tasks and under distribution change.

Self-supervision uses large corpora without labeling every example, but “unlabeled” does not mean uncurated, permission-free, or costless. Collection, deduplication, filtering, and documentation are work. The pretraining task may also favor capabilities different from the final one. Benefits are measured by comparing training from scratch and pretraining with the same labeled data, budget, and protocol.

Reinforcement: a reward defines behavior, not intelligence

In reinforcement learning, an agent observes a state, selects an action, receives a reward, and affects later states. It has to balance exploring alternatives with exploiting what appears to work. Outcomes depend on the environment, reward, horizon, observability, and ability to reset; “learns by itself” hides those design decisions.

Deep Q-Network combined Q-learning and convolutional networks to learn policies in Atari games from pixels and scores. AlphaGo combined policy and value networks, tree search, game records, and self-play. They were different systems inside environments with defined actions and objectives, not stages of one general agent.

An incomplete reward can produce a strategy that maximizes the number while violating its intent. Tests should therefore include shortcuts, side effects, rule changes, and the ability to stop. In a real product, permissions, supervision, and the cost of an action matter as much as simulated return. Sequential success in a game does not establish safe driving or clinical judgment.

Data, hardware, and evaluation determine the outcome

GPUs accelerated parallel matrix operations; specialized accelerators adapt memory, precision, and dataflow to particular workloads. The paper on the first-generation datacenter TPU evaluated defined inference workloads and performance per watt. It does not establish that every ASIC trains faster or uses less energy: the model, batch, precision, utilization, comparison system, and full cycle need to be stated.

Robustness does not follow from average accuracy either. Intriguing properties of neural networks documented that small adversarially constructed perturbations could change network predictions. Responsible testing includes natural and deliberate variations, subgroups, calibration, abstention, and degradation when a sensor or language changes.

An experiment record has six fields: learning signal; model and inductive bias; data and split; objective; resources; and evaluation with failures. It then explains the surrounding system: rules, search, tools, and review. An improvement is credited to the tested difference rather than to “ML” or “DL” as a whole.

To read a demonstration, reconstruct its comparison in practice. Did only the model change, or did data and computation also increase? Does the metric reflect the real costs of false positives and false negatives? Was the test set fixed before tuning? Are variation across runs and subgroups reported? A single number cannot attribute an improvement. The useful result is conditional: under these data, resources, threshold, and population, this system performed this way against this alternative.

Machine learning and deep learning can be parts of very broad systems, but they are neither sufficient nor necessarily exclusive components of a future AGI. The transferable skill is classifying any claim by signal, objective, adaptation, metric, and domain of validity. That map supports choosing a technique today and recognizing its limits without turning learning from data into a countdown to AGI.