Artificial Neural Network, often abbreviated as NN.


A component that can execute computations.


A component of nGraph that acts as a backend for a framework, allowing the framework to define and execute computations.


A builder is a function that creates a sub-graph and returns a root node to the bridge. Fused ops are preferred over builders See also: Basic concepts.

data-flow graph

Data-flow graphs are used to implement deep learning models. In a data-flow graph, nodes represent operations on data and edges represent data flowing between those operations.

dynamic tensor

A tensor whose shape can change from one “iteration” to the next. When created, a framework bridge might supply only partial shape information: it might be all the tensor dimensions, some of the tensor dimensions, or none of the tensor dimensions; furthermore, the rank of the tensor may be left unspecified.


The serialized version of a trained model that can be passed to one of the nGraph backends for computation.


Frameworks provide expressive user-facing APIs for constructing, training, validating, and deploying DL/ML models: TensorFlow*, PaddlePaddle*, MXNet*, PyTorch*, and Caffe* are all examples of well-known frameworks.

function graph

The nGraph Library uses a function graph to represent an op’s parameters and results.


Fusion is the fusing, combining, merging, collapsing, or refactoring of a graph’s functional operations (ops) into one or more of nGraph’s core ops.


An acronym for “Instruction Set Architecture,” an ISA is machine code that is compatible with the underlying silicon architecture. A realization of an ISA is called an implementation. An ISA permits multiple implementations that may vary in performance, physical size, memory use or reuse, and monetary cost among other things. An ISA defines everything a machine-language programmer needs to know in order to program a particular backend device. What an ISA defines will differ among ISAs; in general, it defines things like:

  • supported data types;

  • physical states available, such as the main memory and registers;

  • semantics, such as the memory consistency and addressing modes;

  • low-level machine instructions that comprise a machine language;

  • and the input/output model.

Be careful to not confuse ISAs with microarchitectures.


LSTM is an acronym for “Long Short-Term Memory”. LSTMs extend on the traditional RNN by providing a number of ways to “forget” the memory of the previous time step via a set of learnable gates. These gates help avoid the problem of exploding or vanishing gradients that occur in the traditional RNN.

model description

A description of a program’s fundamental operations that are used by a framework to generate inputs for computation.


NN is an acronym for “Neural Network”. NN models are used to simulate possible combinations of binary logic processing and multi-layer (multi-dimensional) paths through which a data-flow graph may be mapped or computed. A NN does not have centralized storage; rather, a NN manifests as information stored as patterns throughout the network structure. NNs may be Recurrent (feedback loop) or Nonrecurrent (feed-forward) with regard to the network vector.


An op represents an operation. Ops are stateless and have zero or more inputs and zero or more outputs. Some ops have additional constant attributes. Every output of an op corresponds to a tensor and has an element type and a shape. The element types and shapes of the outputs of an op are determined by the inputs and attributes of the op.


In the context of a function graph, a “parameter” refers to what “stands in” for an argument in an op definition.


The term provenance refers to the matching of device code to framework sub-graphs; it is analogous to source code locators in conventional compilers, which associate regions of object code with source files and line numbers.


Quantization refers to the conversion of numerical data into a lower-precision representation. Quantization is often used in deep learning to reduce the time and energy needed to perform computations by reducing the size of data transfers and the number of steps needed to perform a computation. This improvement in speed and energy usage comes at a cost in terms of numerical accuracy, but deep learning models are often able to function well in spite of this reduced accuracy.


Recurrent Artificial Neural Network, often abbreviated as RNN.


In the context of a function graph, the term “result” refers to what stands in for the returned value.


A Recurrent Neural Network is a variety of NN where output nodes from a layer on a data-flow graph have loopback to nodes that comprise an earlier layer. Since the RNN has no “centralized” storage, this loopback is the means by which the ANN can “learn” or be trained. There are several sub-categories of RNNs. The traditional RNN looks like:

\(s_t = tanh(dot(W,x_{t-1}) + dot(U, s_{t-1})\)

where \(x\) is the input data, \(s\) is the memory, and output is \(o_t = softmax(dot(V, s_t))\). Tanh, Dot, and Softmax are all nGraph core Ops.


Stochastic Gradient Descent, also known as incremental gradient descent, is an iterative method for optimizing a differentiable objective function.


The shape of a tensor is a tuple of non-negative integers that represents an exclusive upper bound for coordinate values.

shape propagation

The static process by which assignment of every tensor (or, equivalently, every node output) in the graph is assigned complete shape information.

shared pointer

The C++ standard template library has the template std::shared_ptr<X>. A shared pointer is used like an X* pointer, but maintains a reference count to the underlying object. Each new shared pointer to the object increases the count. When a shared pointer goes out of scope, the reference count is decremented, and, when the count reaches 0, the underlying object is deleted. The function template std::make_shared<X>(...) can be used similarly to new X(...), except it returns a std::shared_ptr<X> instead of an X*.

If there is a chain of shared pointers from an object back to itself, every object in the chain is referenced, so the reference counts will never reach 0 and the objects will never be deleted.

If a referenced b and b wanted to track all references to itself and shared pointers were used both directions, there would be a chain of pointers form a to itself. We avoid this by using shared pointers in only one direction, and raw pointers for the inverse direction. std::enabled_shared_from_this is a class template that defines a method shared_from_this that provides a shared pointer from a raw pointer.

nGraph makes use of shared pointers for objects whose lifetime is hard to determine when they are allocated.


An abstract “action” that produces zero or more tensor outputs from zero or more tensor inputs. Steps correspond to ops that connect nodes.


Tensors are maps from coordinates to scalar values, all of the same type, called the element type of the tensor.


The interface backends implement for tensor use. When there are no more references to the tensor view, it will be freed when convenient for the backend.


To provide optimizations with nGraph, we first confirm that a given workload is “validated” as being functional; that is, we can successfully load its serialized graph as an nGraph function graph


A workload is an Artificial Intelligence training job that can be executed via one of the distributed training methods.