Semi-supervised learning (SLA)

What is semi-supervised learning?

An algorithm is trained with both labelled and unlabelled data. Semi-supervised learning thus enables time- and cost-efficient learning. In the field of Artificial intelligence a learning process is needed that allows the system to intelligently learn relationships. Unlike supervised learning, semi-supervised learning is able to classify data quickly and effectively as in unsupervised learning.

A variety of scenarios are possible where data with labels is actually not readily available. For example, semi-supervised learning can achieve optimal results with a fraction of labelled data, such as hundreds of training examples. Semi-supervised learning can handle those types of datasets that choose either supervised learning or unsupervised learning - without having to compromise.

When is semi-supervised learning used?

semi-supervised or semi-supervised learning involves feature estimation of appropriately labelled and unlabelled data. With this approach, not as much labelled data is needed, which is often relatively expensive to create. Unlabelled data is much cheaper and this can also be used for learning. The challenge is in compiling this Training datato provide a ratio of labelled and unlabelled data of high overall significance to the algorithm.

The aim is to assign a correct label to the unlabelled data. This can be achieved with so-called label propagation. Such a method has similarities to a Cluster analysis on. The data can be divided into clusters and then within the cluster the unlabelled data can easily be assigned the same labels.

What is label spreading?

Label spreading is a form of semi-supervised learning algorithm. This algorithm by Dengyong Zhou et al. appeared in their article entitled "Learning with Local and with Global Consistency" in 2003. Thus, the intuition for a broader approach to semi-supervised learning is that nearby points in the input space should have the same label and the points in the same structure or manifold in the input space should have the same label.

Label propagation is practically borrowed from a technique in experimental psychology called a propagation activation network. Thus, points in the data set are connected quite simply based on the relative distances in the input space in such a graph. Symmetrically normalised is the weight matrix of this graph, similar to spectral clustering. The information is then passed through the graph, which is adjusted to capture the structure in the input space. Thus, finally, the label of each unlabelled point is set to the very class where it just got the most information during the iteration process. The use of label spreading helps to save costs.


What are hyperparameters?

A hyperparameter is a certain parameter with which the learning process can be controlled. They can be used for model selection tasks - on the other hand, there are also algorithm hyperparameters which in principle have no influence on the performance of the model, but which influence the speed and quality of the learning process. Thus, one can speak of model hyperparameters that influence the topology and the size of Neural networks characterise. On the other hand, algorithm hyperparameters are the learning rate and the mini-batch size.

Different model-training algorithms require different hyperparameters - some simple Algorithms do not require any. With these hyperparameters, the training algorithm learns the parameters from given data. The time needed to train and test a model may depend on what choice is made in hyperparameters. A hyperparameter can usually be a continuous or an integer type. This can lead to mixed type optimisation problems. The existence of some hyperparameters is conditional on the values of others, such as the size of each hidden layer in a neural network, which may be conditional on the number of layers.

Hyperparameter optimisation is used to search for optimal hyperparameters in machine learning. A hyperparameter is determined before the actual model training.

When is hyperparameter optimisation used?

Automated machine learning and deep neural networks use hyperparameter optimisation. There are black-box function optimisation methods based on model-free methods and Bayesian optimisations. The optimisations entail high computational requirements. Every machine learning system has hyperparameters. The hyperparameters are automatically set to optimise performance.

What is hyperparameter optimisation?

Hyperparameters are adjustable parameters that you can use to control the model training process. In neural networks, you can set the number of hidden layers and determine the number of nodes that are on each layer. The performance of the model depends heavily on the hyperparameters.

Hyperparameter optimisation is performed with the process of searching for a suitable configuration. The configuration selected is the one that delivers the best performance. Such a process is practically always computationally intensive and must be carried out manually.

What are the differences between parameters and hyperparameters?

A model parameter is a configuration variable that is internal to the model used and whose value is estimated from the data. Such parameters are needed to make predictions. The values define the capability of the model for your problem. They are estimated or learned from the data. Usually they are not set manually by the user. Model parameters are often stored as part of the learned model. They are the key to machine learning.

In contrast, a hyperparameter is a configuration that is external to a model and whose value cannot be estimated from the data. These hyperparameters are often used to estimate model parameters in processes.


What is heuristics?

Heuristics refers to an analytical procedure of arriving at conclusions and solutions with limited knowledge. The inferred statements may deviate from the optimal solution. The comparison between the optimal solution and the solution of the heuristic procedure determines the quality of the heuristic procedures.

The best-known heuristic methods include trial and error algorithms. They also include statistical evaluations of certain random samples and exclusion procedures. Such procedures are based on experience already gained.

Through heuristics it is possible that a Optimum solution in a short time can be found. In this way, shortened computational paths are used to perform calculations more quickly. This has made it possible for programmes like Deep Blue or AlphaGo be able to beat leading professional players in chess and Go respectively.

Areas of application for heuristics can be found in speech and face recognition, in character recognition (OCR), in the Data mining and in general knowledge-based systems.

How is a heuristic procedure defined?

In the more recent philosophy of science, heuristics are as an assessment criterion for theories, as well as for entire science programmes (paradigm) and are of particular importance. However, it is not exclusively the information content that is evaluated, but above all the inherent potential for further development of the state of knowledge.

Heuristic methods can provide a procedure for solving general problems. Especially when no clear solution strategy is known, they can be used successfully. They are primarily based on subjective experience. Traditional behavioural patterns can also play a role.

They can be used especially for problem areas that are poorly structured or difficult to grasp. In this way mathematical problems elegantly solved and their time required considerably reduced can be used. Such solution methods can be used without proof of convergence for problems for which no converging methods exist.

What types of heuristics are there?

Numerous different methods are used in science. Their structure usually depends on the area of application in which they are used.

  • Take-the-Best heuristic procedures exploit the best possible strategy.
  • Recognition heuristics are based on the recognition effect.
  • Availability heuristics are based on individual memories.
  • Representational heuristics have a prototype and allow inferences to be made in the case of similarities.
  • Anchor heuristics can possibly benefit from the anchor effect, whereby arbitrarily chosen anchors influence people in their decision-making process.
  • The expert heuristic can use expert knowledge.
  • Sympathy heuristic evaluates statements and appearance of persons.
  • The consensus heuristic pays attention to majority opinions.
  • In Operations Research, heuristics exist as opening procedures, as improvement procedures, as incompletely exact procedures and as compound procedures. A distinction is made between problem-specific and universal heuristics.
  • There are also metaheuristics, such as the ant algorithm, different types of evolutionary algorithms, or simulated annealing, tabu search and also variable neighbourhood search.


What is the hypothesis in Artificial Intelligence?

A Hypothesis is an inherently logical assumption made on the basis of assumptions and experience and not yet proven is.

For the Artificial Intelligence (AI) there are AI hypotheses in two different schools of thought. With the Strong AI hypothesis the machine is able to actually think like a human being and develop real self-awareness. On the other hand, the Weak AI hypothesis the machine is only able to take over individual intelligent functions as a substitute for humans. For example, there are pattern recognition, search programmes or even systems (such as cars, computer programmes) that optimally adapt their behaviour to the operator (as in the case of sporty drivers or also preferred functions).

Since little is generally known yet about how human intelligence can be described holistically, a useful benchmark for a strong AI hypothesis is lacking. The hardware of the brain is not fundamentally different from that of a computer. Each element, whether it is a transistor or a nerve cell, performs very simple functions - these can be replicated by computers.

Technological singularity impressively describes a widespread hypothesis according to which the development of an artificial superintelligence is possible, which can enable rapid technological growth. Thus, a powerful AI could evolve with unstoppable sequences of self-improvement cycles, resulting in an "intelligence explosion". Such a superintelligence could far surpass human intelligence.

Which applications with hypothesis generation are known?

There is AI in healthcare that can be used successfully. One of the best known is IBM Watson. This system understands natural language and it can respond to questions that are asked. Such a system evaluates patient data and can form hypotheses based on this.

There are also AI applications such as virtual online health assistants and functional chatbotswhich help patients and clients in the health sector to find medical information directly. Appointments can also be made. AI applications can help understand billing processes and handle various administrative procedures.

Can robust hypotheses be created through AI?

The human brain has the ability to build mental models from experience and thus to systematically draw broader conclusions beyond current sensory impressions, which even make explanations possible. There is the form of explanatory structure that can also be called a generative model. This is based on a corresponding multitude of different experiences and generates a hypothesis through certain sensory perceptions. In this way, statements can be made about how objects move or whether they are hidden. These hypotheses are significantly more robust than classification-oriented systems, which are used in the Deep Learning underlying.


What is Human-in-the-Loop?

A human-in-the-loop is a human who trains, tests and optimises an AI system to achieve more reliable results. The AI, like a normal student, makes mistakes or gets certain details wrong when starting a new activity. For example, a system can be taught to detect animals in the sea and it can distinguish an octopus with its particular shape from other animals.

However, difficulties may arise if other fish have a similar shape and colour. In such cases, there is the possibility to intervene with a human-in-the-loop and enter different characteristics in the system to search for. In such a way, the system can arrive at more accurate answers. A major advantage of human-instruction input is that two different types of intelligence can be used virtually simultaneously. This way, data can be provided and the AI system can check and evaluate its progress.

Humans can contribute their own knowledge, with which they have learned themselves, and can combine this with the speed of the computer. Thus, there is a fantastically large potential with this artificial intelligence and through dynamic cooperation, disadvantages of humans and machines can easily be compensated and thus more accurate results can be comfortably achieved.

What human-in-the-loop simulators are there?

There are quite a few different simulators for this concept:

  • Flight simulators
  • Vehicle simulators
  • Marine simulators
  • Most diverse video games
  • Supply Chain Management Simulators
  • Digital puppetry

When is a human involved in the calculation process?

Human-in-the-loop systems work together with a human supervisor. This supervisor can help in critical situations. The system usually implements what it has learned on its own. In this way, many processes are automated. However, in difficult and unknown situations, a human can be asked to make difficult decisions or to explain a new situation to the machine.

How does Human-in-the-Loop work?

Human-in-the-loop combined Supervised Learning with Active Learning and is an essential component of AI applications. Predictions should be made as accurately as possible. The Human in the Loop system provides efficient and fast procedures for model training and prediction. It is a useful method for identifying informative examples. There are ergonomic tools for labelling appropriate training data.