Multi-Input Models

A multi-input or multi-output model is easier to understand when you stop imagining it as one straight stack of layers. A Keras functional model can receive several streams of data, process each stream through its own branch, and merge the resulting representations. It can also share one representation and branch into several named prediction heads. The graph therefore has places where information is separate and places where it is shared.

A single-input model receives one stream of data. A multi-input model receives several streams and gives each stream its own path before combining the resulting representations. For example, a question-answering system can receive a question through one input and a text snippet through another. The question and text travel independently until a merge operation brings their branch representations together.

The main construction stages of a two-input functional model are input layers, independent branch encoders, a merge operation, and an output layer. Input names such as text and question identify the input nodes. Those names are useful later because dictionary-based training uses them to associate each array with the correct branch. The model construction is explicit: it receives both input tensors and produces the answer tensor.

The branches meet through a Keras merge operation such as concatenate or add. The merge line is not a minor implementation detail. It determines how the branch representations are brought together, so it is one of the first places to inspect when the model behaves unexpectedly.

Training data must preserve the model's input structure. With list-based training, input arrays are supplied in the same order as the model's inputs. This form is concise, but the order must remain correct. With dictionary-based training, arrays are supplied under the names of the input nodes. This makes the pairing explicit and requires named inputs.

A multi-output model makes several predictions from the same input or shared internal representation. Each output can represent a different task, so each head needs a loss suited to what it predicts. The source example uses an age head for scalar regression, an income head for classification over income groups, and a gender head for binary classification. These tasks therefore use different kinds of losses rather than one loss reused blindly for every output.

During training, one target array is paired with each output. Keras accepts those targets as a list ordered like the model outputs or as a dictionary keyed by output names. The dictionary form makes the pairing explicit; the list form is shorter but depends on keeping the output order correct.

The separate output losses do not remain isolated during optimization. Keras combines the individual losses into one global loss used to optimize the model. Because the output heads share a representation, the combined result determines how the shared part of the network is trained.

A plain combination does not guarantee that every task influences learning equally. If one individual loss is numerically much larger than the others, optimization can be driven mainly by that task. Loss weights change each loss's contribution before the global loss is formed, helping balance outputs whose losses have different numerical scales.

A multi-input model can fail at several distinct boundaries. Treat it as a trace: data enters an input node, passes through its branch, reaches the merge operation, and contributes to the output. For a multi-output model, trace each prediction head to its loss, optional weight, and target data. The most useful debugging question is where the actual path first differs from the intended graph.