Abstract: Techniques for implementing unsupervised anomaly detection by self-prediction are provided. In one set of embodiments, a computer system can receive an unlabeled training data set comprising a plurality of unlabeled data instances, where each unlabeled data instance includes values for a plurality of features. The computer system can further train, for each feature in the plurality of features, a supervised machine learning (ML) model using a labeled training data set derived from the unlabeled training data set, receive a query data instance, and generate a self-prediction vector using at least a portion of the trained supervised ML models and the query data instance, where the self-prediction vector indicates what the query data instance should look like if it were normal. The computer system can then generate an anomaly score for the query data instance based on the self-prediction vector and the query data instance.

Abstract: Some embodiments provide a method for performing data message processing at a smart NIC of a computer that executes a software forwarding element (SFE). The method determines whether a received data message matches an entry in a data message classification cache stored on the smart NIC based on data message classification results of the SFE. When the data message matches an entry, the method determines whether the matched entry is valid by comparing a timestamp of the entry to a set of rules stored on the smart NIC. When the matched entry is valid, the method processes the data message according to the matched entry without providing the data message to the SFE executing on the computer.

Abstract: Some embodiments provide a method for, at a network interface controller (NIC) of a computer, accessing data in a network. From the computer, the method receives a request to access data stored at a logical memory address. The method translates the logical memory address into a memory address of a particular network device storing the requested data. The method sends a data message to the particular network device to retrieve the requested data.

Abstract: Some embodiments provide a method for performing data message processing at a smart NIC of a computer that executes a software forwarding element (SFE). The method stores (i) a set of cache entries that the smart NIC uses to process a set of received data messages without providing the data messages to the SFE and (ii) rule updates used by the smart NIC to validate the cache entries. After a period of time, the method determines that the rule updates are incorporated into a data message processing structure of the SFE. Upon incorporating the rule updates, the method deletes from the smart NIC (i) the rule updates and (ii) at least a subset of the cache entries.

Abstract: Some embodiments provide a method for performing data message processing at a smart NIC of a computer that executes a software forwarding element (SFE). The method stores (i) a set of cache entries that the smart NIC uses to process a set of received data messages without providing the data messages to the SFE and (ii) rule updates used by the smart NIC to validate the cache entries. After a period of time, the method determines that the rule updates are incorporated into a data message processing structure of the SFE. Upon incorporating the rule updates, the method deletes from the smart NIC (i) the rule updates and (ii) at least a subset of the cache entries.

Abstract: Some embodiments provide a method for transmitting data at a network interface controller (NIC) of a computer that operates as a server. The computer includes multiple storage devices. The method receives a request from a client device for a particular file. The method translates the particular file into a memory location corresponding to a particular one of the storage devices at the computer. The method transmits the requested file from the particular storage location to the client device.

Abstract: Some embodiments provide a method for performing data message processing at a smart NIC of a computer that executes a software forwarding element (SFE). The method determines whether a received data message matches an entry in a data message classification cache stored on the smart NIC based on data message classification results of the SFE. When the data message matches an entry, the method determines whether the matched entry is valid by comparing a timestamp of the entry to a set of rules stored on the smart NIC. When the matched entry is valid, the method processes the data message according to the matched entry without providing the data message to the SFE executing on the computer.

Abstract: Some embodiments provide a method for, at a network interface controller (NIC) of a computer, accessing data in a network. From the computer, the method receives a request to access data stored at a logical memory address. The method translates the logical memory address into a memory address of a particular network device storing the requested data. The method sends a data message to the particular network device to retrieve the requested data.

Abstract: Some embodiments provide a method for sending data messages at a network interface controller (NIC) of a computer. From a network stack executing on the computer, the method receives (i) a header for a data message to send and (ii) a logical memory address of a payload for the data message. The method translates the logical memory address into a memory address for accessing a particular one of multiple devices connected to the computer. The method reads payload data from the memory address of the particular device. The method sends the data message with the header received from the network stack and the payload data read from the particular device.

Abstract: Some embodiments provide a method for performing distributed machine learning (ML) across multiple computers. At a smart network interface controller (NIC) of a first computer, the method receives a set of ML parameters from the first computer related to training an ML model. The method compresses the set of ML parameters based on a current state of a connection to a central computer that receives sets of ML parameters from a plurality of the computers. The method sends the compressed set of ML parameters to the central computer for the central computer to process the compressed set of ML parameters along with corresponding sets of ML parameters received from the other computers of the plurality of computers.

Abstract: Some embodiments provide a method for using a machine learning (ML) model to respond to a query, at a smart NIC of a computer. The method receives a query including an input. The method applies a first ML model to the input to generate an output and a confidence measure for the output. When the confidence measure for the output is below a threshold, the method discards the output and provides the query to the computer for the computer to apply a second ML model to the input.

Abstract: Techniques for implementing intelligent data partitioning for a distributed machine learning (ML) system are provided. In one set of embodiments, a computer system implementing a data partition module can receive a training data instance for a ML task and identify, using a clustering algorithm, a cluster to which the training data instance belongs, the cluster being one of a plurality of clusters determined via the clustering algorithm that partition a data space of the ML task. The computer system can then transmit the training data instance to a ML worker of the distributed ML system that is assigned to the cluster, where the ML worker is configured to build or update a ML model using the training data instance.

Abstract: Techniques for implementing a tree-based ensemble classifier comprising an internal load balancer are provided. In one set of embodiments, the tree-based ensemble classifier can receive a query data instance and select, via the internal load balancer, a subset of its decision trees for processing the query data instance. The tree-based ensemble classifier can then query each decision tree in the selected subset with the query data instance, combine the per-tree classifications generated by the subset trees to generate a subset classification, and determine whether a confidence level associated with the subset classification is sufficiently high. If the answer is yes, the tree-based ensemble classifier can output the subset classification as a final classification result for the query data instance. If the answer is no, the tree-based ensemble classifier can repeat the foregoing steps until a sufficient confidence level is reached or until all of its decision trees have been selected and queried.

Abstract: Techniques for implementing intelligent data partitioning for a distributed machine learning (ML) system are provided. In one set of embodiments, a computer system implementing a data partition module can receive a training data instance for a ML task and identify, using a clustering algorithm, a cluster to which the training data instance belongs, the cluster being one of a plurality of clusters determined via the clustering algorithm that partition a data space of the ML task. The computer system can then transmit the training data instance to a ML worker of the distributed ML system that is assigned to the cluster, where the ML worker is configured to build or update a ML model using the training data instance.

Abstract: Techniques for accelerating the training of machine learning (ML) models in the presence of network bandwidth constraints via data instance compression. For example, consider a scenario in which (1) a first computer system is configured to train a ML model on a training dataset that is stored on a second computer system remote from the first computer system, and (2) one or more network bandwidth constraints place a cap on the amount of data that may be transmitted between the two computer systems per training iteration. In this and other similar scenarios, the techniques of the present disclosure enable the second computer system to send, according to one of several schemes, a batch of compressed data instances to the first computer system at each training iteration, such that the data size of the batch is less than or equal to the data cap.

Abstract: Techniques for quantizing training data sets using machine learning (ML) model metadata are provided. In one set of embodiments, a computer system can receive a training data set comprising a plurality of features and a plurality of data instances, where each data instance includes a feature value for each of the plurality of features. The computer system can further train a machine learning (ML) model using the training data set, where the training results in a trained version of the ML model, and can extract metadata from the trained version of the ML model pertaining to the plurality of features. The computer system can then quantize the plurality of data instances based on the extracted metadata, the quantizing resulting in a quantized version of the training data set.

Abstract: Techniques for implementing importance sampling via machine learning (ML)-based gradient approximation are provided. In one set of embodiments, these techniques include (1) training a deep neural network (DNN) on a training dataset using stochastic gradient descent and (2) in parallel with (1), training a separate ML model (i.e., gradient approximation model) that is designed to predict gradient norms (or gradients) for the data instances in the training dataset. The techniques further include (3) applying the gradient approximation model to the training dataset on a periodic basis to generate gradient norm/gradient predictions for the data instances in the training dataset and (4) using the gradient norm/gradient predictions to update sampling probabilities for the data instances. The updated sampling probabilities can then be accessed during the ongoing training of the DNN (i.e., step (1)) to perform importance sampling of data instances and thereby accelerate the training procedure.

Abstract: Techniques for implementing efficient federated learning of deep neural networks (DNNs) using approximation layers are provided. In one set of embodiments, given a DNN M with k original layers {L1, . . . , Lk}, k approximation layers {L1?, . . . , Lk?} can be created that correspond (i.e., map) to the k original layers. Each approximation layer can have the same number of inputs and outputs as its corresponding original layer, but can be smaller in size (i.e., have fewer parameters). Then, at the time of training DNN M via federated learning, for each participating client c during a training round r, a parameter server can transmit, for i=1, k, either (1) the current parameter values for approximation layer Li? alone, or (2) the current parameter values for both original layer Li and approximation layer Li? to client c. In response, client c can train its local copy of DNN M in accordance with the received parameter values.

Abstract: Techniques for implementing unsupervised anomaly detection via supervised methods are provided. In one set of embodiments, a computer system can train an unsupervised anomaly detection classifier using an unlabeled training data set and classify the unlabeled training data set via the trained version of the unsupervised classifier, where the classifying generates anomaly scores for the data instances in the unlabeled training data set. The computer system can further construct a labeled training data set that includes a first subset of data instances from the unlabeled training data set whose anomaly scores are below a first threshold and a second subset of data instances from the unlabeled training data set whose anomaly scores are above a second threshold. The computer system can then train a supervised anomaly detection classifier using the labeled training data set.

Abstract: Techniques for performing predictability-driven compression of training data sets used for machine learning (ML) are provided. In one set of embodiments, a computer system can receive a training data set comprising a plurality of data instances and can train an ML model using the plurality of data instances, the training resulting in a trained version of the ML model. The computer system can further generate prediction metadata for each data instance in the plurality of data instances using the trained version of the ML model and can compute a predictability measure for each data instance based on the prediction metadata, the predictability measure indicating a training value of the data instance. The computer system can then filter one or more data instances from the plurality of data instances based on the computed predictability measures, the filtering resulting in a compressed version of the training data set.