A general framework for adaptive processing of data structures

By Paolo Frasconi, Marco Gori, Alessandro Sperduti. 1998

Subfields: Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Representation & Self-Supervised Learning

Citations in the corpus

Graph Attention Networks (2017, methodology): “Early work used recursive neural networks to process data represented in graph domains as directed acyclic graphs (Frasconi et al., 1998; Sperduti & Starita, 1997).”
Neural Turing Machines (2014, background): “Considering this property, we do not feel that it is urgent or even necessarily valuable to build explicit parse trees to merge composite structures greedily (Pollack, 1990) (Socher et al., 2012) (Frasconi et al., 1998).”
The Graph Neural Network Model (2009, neutral): “Finally, it can be observed that the total average error for FNNs is about 50% larger than the GNN error (12.7 for nonlinear GNNs, 13.5 for linear GNNs, and 22.8 for FNNs).”
Unsupervised Inductive Whole-Graph Embedding by Preserving Graph Proximity (2019, background): “• Supersource: Dated back to early works (Frasconi et al., 1998; Scarselli et al., 2009), the idea is to introduce a “dummy/super node” connecting to every node in the graph, so that during aggregation, this supersource node absorbs information from all the nodes.”
Practical Recommendations for Gradient-Based Training of Deep Architectures (2012, methodology): “This idea has been generalized from sequences and 2-D images to arbitrary graphs with recursive neural networks or recursive graphical models (Pollack, 1990; Frasconi et al., 1998; Bottou, 2011; Socher et al., 2011), Markov Logic Networks (Richardson and Domingos, 2006) and relational learning (Geto”
Image Generation from Scene Graphs (2018, methodology): “More closely related to our work are Graph Neural Networks (GNNs) [11, 13, 46] which generalize recursive neural networks [8, 49, 48] to operate on arbitrary graphs.”
Exploiting the past and the future in protein secondary structure prediction (1999, methodology): “Because the unrolled network is acyclic, the generalized backpropagation algorithm can be derived as a special case of the backpropagation through structure algorithm (Frasconi et al., 1998).”
Architectural Bias in Recurrent Neural Networks: Fractal Analysis (2002, background): “It should be mentioned that RNNs can be canonically generalized to so-called recursive networks for processing tree structures and directed acyclic graphs, hence providing a very attractive connectionist framework for symbolic or structural data (Frasconi, Gori & Sperduti, 1998; Hammer, 2002).”
Deep Learning on Graphs: A Survey (2018, unknown): “Motivated by recursive neural networks [33], a recursive deﬁnition of states is adopted [23]: where Fð(cid:7)Þ is a parametric function to be learned.”
A Gentle Introduction to Deep Learning for Graphs (2019, background): “as a purely theoretical model [41] and later more practically through efficient approximated distributions [6].”
Understanding over-squashing and bottlenecks on graphs via curvature (2021, background): “…few years, deep learning on graphs and in particular graph neural networks (GNNs) (Sperduti, 1994; Goller & Kuchler, 1996; Sperduti & Starita, 1997; Frasconi et al., 1998; Gori et al., 2005; Scarselli et al., 2008; Bruna et al., 2014; Defferrard et al., 2016; Kipf & Welling, 2017; Gilmer et al.,…”
Sequence Learning (2001, methodology): “[24], where graphical models are used to describe in a unified framework both neural and probabilistic (Bayesian Networks) transductions involving data structures.”
Learning Nonsparse Kernels by Self-Organizing Maps for Structured Data (2009, methodology): “Recent developments in machine learning have produced methods capable of processing structured information directly, such as kernel methods [8] and artificial neural networks (NNs) [2], [ 9 ].”
A Comparison between Recursive Neural Networks and Graph Neural Networks (2006, background): “Recursive Neural Networks Recursive Neural Networks (RNNs) were originally proposed in [15], [16], [20].”
Prediction of contact maps by GIOHMMs and recurrent neural networks using lateral propagation from all four cardinal corners (2002, background): “…other general set of functions) which S64 amounts to consider that the state of each node is a deterministic vector that depends on the value of the parent vectors through a function that is implemented by a neural network (Baldi and Chauvin, 1996; Frasconi et al., 1998; Pollastri et al., 2001a,b).”
On the Complexity of Neural Network Classifiers: A Comparison Between Shallow and Deep Architectures (2014, unknown): “In addition, even recurrent/recursive models [13], [14] and graph neural networks [15], which cannot be strictly deﬁned deep architectures, actually implement the unfolding mechanism during learning, which, in practice, produces networks that are as deep as the data they have to model.”